Underwater Image Enhancement via Multi-Scale Feature Fusion Network Guided by Medium Transmission

Underwater Image Enhancement via Multi-Scale Feature Fusion Network Guided by Medium Transmission | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Underwater Image Enhancement via Multi-Scale Feature Fusion Network Guided by Medium Transmission Hao Yang, Hongqin Cai, Chenxu Jiang, Ruiteng Zhang, Jian Zhang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4082073/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Due to the complexity of underwater imaging environments, images captured via optical vision systems often exhibit significant degradation. To combat this issue, we introduce a multi-scale feature fusion underwater image enhancement network, termed MFUNet. MFUNet is a novel multi-scale feature fusion network, guided by medium transmission, ensures the content integrity of the reconstructed image by leveraging interaction features among non-adjacent layers. This approach addresses the common problem of the loss of image detail features. Moreover, MFUNet enhances the response to high-frequency information by employing edge loss, thereby improving sensitivity to edges and textures. By deepening the network hierarchy, the image undergoes deep encoding and decoding, which maximizes the multi-color space encoder's and multi-scale feature fusion's potential in color representation and enhances the structural similarity and overall quality of the image. It is worth noting that we achieved superior performance by utilizing fewer model parameters. Extensive experiments across various datasets demonstrate that our method surpasses comparative methods in both visual quality and quantitative metrics. Physical sciences/Mathematics and computing/Computational science Physical sciences/Mathematics and computing/Computer science Physical sciences/Mathematics and computing/Scientific data Image enhancement Underwater imaging Deep learning Multi-scale feature fusion Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 1. Introduction Enhancing underwater images is a critical task for advancing marine applications and services, including underwater surveillance, image and video compression and transmission, and target detection. However, underwater imagery often suffers from significant color distortion, and the visuals are marred by low clarity and contrast due to the wavelength-dependent absorption of light, as well as scattering and refraction. Consequently, Improving the visual quality of underwater images through image enhancement is a formidable challenge. Traditional underwater image enhancement methods mainly include non-physical model-based methods 1–4 and physical model-based methods 5–9 . Physical model-based methods 9–15 focus on the accurate estimation of medium transmission and use the estimated medium transmittance and key underwater imaging parameters, such as uniform background light, to recover underwater images to obtain high quality images. They 16 proposed a Dark Channel Prior method for image dehazing, called DCP. The method obtains clear images by estimating atmospheric light and transmission maps. Due to its simplicity and effectiveness, many recovery-based UIE methods have modified the DCP to address the severe attenuation of red light in water. Chiang and Chen 10 used a defogging algorithm that compensates for light attenuation to enhance underwater images. Galdran et al. 17 proposed a Red-Channel Prior method (RCP) to recover the relevant colors at short wavelengths in order to restore contrast. Based on DCP, Peng et al. 18 proposed an Underwater DCP (UDCP) that only considers green and blue channels for underwater image restoration. The non-physical model method achieves the purpose of restoration by achieving contrast enhancement. They 2 proposed a method based on a non-physical model, called Histogram Equalization (HE), which uses pixel-value transformations to transform the original image into roughly the same number of pixels at most grey levels to recover the image. The limitations of traditional image methods mainly lie in ignoring the underwater imaging mechanism, resulting in insufficient or excessive enhancement, such as HE 2 , or being time-consuming or sensitive to the diversity of underwater scenes, such as DCP 16 , RCP 17 etc. With the development of deep learning technology, image enhancement methods based on deep learning are gradually becoming mainstream 19–21 . The deep learning approach employs hierarchical feature extraction and original image reconstruction to produce high-quality restored images.The multi-color space encoder proposed by Ucolor 22 can integrate the features of different color spaces into a unified structure, improving the color space expression of images. However, it does not consider the feature fusion between different levels, does not pay attention to the structural similarity and scale difference of the restored image, and does not smooth the edges during the enhancement process, resulting in a large semantic gap between each feature layer and the loss of important details in the enhancement result. The U-Net 23 architecture introduces a fully symmetrical (U-shaped) network structure. It concatenates down-sampled and up-sampled feature maps of identical dimensions through skip connections, enabling the effective integration of high-level and low-level features. This method proves particularly beneficial in scenarios with limited sample sizes, offering rapid processing speeds. However, the U-Net 23 architecture falls short in representing features across different color spaces, and its feature depth is inadequate for leveraging the full potential of multi-scale feature fusion and multi-color space encoders. To address the issues identified, we leverage the strengths of both physical model-based and deep learning-based approaches. By integrating multi-scale fusion with enriched encoder features within the context of target detection tasks, we aim to correct color bias and amplify the contrast of underwater images. In our research, we employ a medium transfer map to quantify the extent of quality degradation in underwater images, facilitating targeted enhancement of deteriorated areas. And we use adaptive spatial fusion algorithms 24 to achieve feature interactions at non-adjacent levels through a progressive feature pyramid network and filter features in the multilevel fusion process so as to avoid large semantic gaps between non-adjacent layers, which allows us to retain useful information for fusion. Different from their method 22 , we use the medium transfer map to guide the feature pyramid, so that degraded regions can be enhanced when different feature layers interact. This method not only incorporates the underwater imaging mechanism into the network, but also enables the network to better retain underwater image details and improve the generalization ability of the network, thereby accelerating network optimization and improving performance. Inspired by their method 25 , we have incorporated an edge penalty term into the training regimen of our network. This adjustment heightens the network's sensitivity to edges and textures in underwater images. Our network is purely data-driven, designed to accommodate inaccuracies in media transfer estimates without compromising performance. In Fig. 1 , we give a representative example, which is a comparison diagram between our proposed MFUNet and the other two underwater image enhancement methods. As shown in the figure, the fusion-based approach 26 (Fig. 1 (b)) fails to repair the degraded regions and ignores them instead; the deep learning-based approach 22 (Fig. 1 (c)) handles the details poorly. Our MFUNet(Fig. 1 (d)) performs better in repairing degraded areas and enhancing details, and the repair results are more comfortable and natural. The main contributions of this article are as follows: Medium Transmission-Guided Multi-Scale Fusion Network: We introduce a medium transmission-guided multi-scale fusion network, where features in the multilevel fusion process undergo adaptive spatial fusion operations. This strategy prevents the loss or degradation of information throughout the multilevel transmission process, ensuring the retention of valuable information. Additionally, by utilizing a medium transmission map, we direct the network's attention towards regions of quality degradation, thereby harmonizing the strengths of both physical model-based and deep learning-based approaches. Edge Loss Function Implementation: Our methodology incorporates an edge loss function by introducing an edge penalty term during the network's training phase. This compels the network to heighten its sensitivity towards the edges and textures of underwater images. Consequently, this enhances the network's responsiveness to high-frequency information, leading to an improvement in image quality. Enhanced Network Encoding and Decoding Levels: By deepening the encoding and decoding levels of our network, we capitalize on the benefits of the adaptive spatial fusion algorithm and the multi-color encoder. This amplifies the network's feature representation capabilities and boosts overall performance. Performance Benchmarking: Our MFUNet achieves unparalleled performance in terms of objective metrics when compared to contemporary benchmark methods. 2. Related Work In this section, we discuss related work on underwater image enhancement. Underwater image enhancement methods can be roughly divided into: methods based on physical models, methods without physical models, and methods based on deep learning. The methods without physical model directly modifies the image pixel values to improve visibility, such as multi-scale fusion 26–28 , variational optimization 29 , and pixel distribution adjustment 30 , etc. For example, Ancuti et al. 26 first obtain the color correction and contrast enhancement maps of an underwater image respectively; then calculate the corresponding weight maps; and finally combine the respective advantages to obtain the enhancement maps. Further, Ancuti et al. found that the information contained in at least one-color channel is completely lost in unfavorable scenes such as nighttime, underwater and uneven artificial illumination. So they proposed a color channel compensation preprocessing method to improve the traditional restoration method. The physical model-free methods ignore the underwater imaging mechanism, which tends to produce excessive or insufficient results and introduces artificial colors that produce unstable enhancement results. The physical model-based approach 12,17,18,31,32 treats single underwater image enhancement as a pathological inverse problem, where the parameters of the underwater image formation model are estimated by manual a priori. The prior includes underwater dark channel prior 31 , red channel prior 17 , minimum informati 12 on prior and fuzzy prior 12 , etc. These methods take into account the scattering and attenuation of light and achieve good results. Akkaynak and Treibitz 33 proposed a model that uses modified underwater images to form a model with physical accuracy. And based on the revised model, a new underwater image color correction method 34 based on RGB-D image pairs was proposed. Peng and Cosman 32 proposed an underwater image depth estimation algorithm based on image blur and light absorption, which uses the estimated depth and underwater imaging model to enhance underwater images. Peng et al. 35 further proposed generalizing the dark channel before processing different images captured under severe weather. However, in challenging underwater conditions, statistical priors may fail, and underwater image formation models may produce unreliable enhancement results 33,36 due to different scene properties. Moreover, this method may require manual correction and is time-consuming. Deep learning-based methods have made significant progress in underwater image enhancement 20–22,37-42 . Due to the lack of underwater images and corresponding clean image pairs, previous works have used generative adversarial networks (Generative Adversarial Network, GAN) to synthesize underwater image pairs or perform solution pair learning. Li et al. 20 first used a Generative Adversarial Network (GAN) to synthesize degraded images and proposed a two-stage refinement network. To meet the requirement for paired training data. 19 proposes a weakly supervised underwater color correction network (UCcycleGAN). Uplavikar et al. 38 introduced a simple classifier to make the GAN model more discriminative for different water types. In order to solve the problem of lack of unpaired data sets in supervised learning 21 , real underwater images were simulated based on the underwater imaging physical model and 10 different water types 43 . And they proposed a lightweight CNN model trained on each of the ten different water types and an underwater image enhancement benchmark based on real-world underwater image pairs respectively (UIEB) 41 . PRWnet further proposes a gate fusion network that fuses three enhanced inputs to obtain enhanced underwater images. 22 proposes a multi-color space embedding network that combines the advantages of physical models to deal with color casts and low contrast in underwater images. Existing underwater image enhancement methods mostly focus on the spatial domain, and it is difficult to recover high and low frequency mixed distortion images, 25 proposes a wavelet enhancement-based learning network to gradually refine underwater images in the spatial and frequency domains. The current deep learning-based methods do not focus on the structural similarity and image quality of the recovered images, ignore the targets with scale differences, and the network hierarchy is shallow, which cannot give full play to the advantages of each module. Compared with existing deep learning-based underwater image enhancement methods, our approach has the following advantages: We introduce an adaptive spatial fusion algorithm and guide it through a medium transmission map, and we integrate the underwater imaging principle into the deep structure so that the restored image can adaptively retain the original image information, continuously enhancing the degraded area. We calculate the edge loss using the Laplace operator and make it serve as a penalty term during the network training process, enabling the network to restore the high-frequency information of the image. We have deepened the depth of the network, so that each module of the network can play a better performance. Our method achieves excellent results on a variety of underwater image datasets. 3. Proposed Method In this section, the proposed MFUNet framework is introduced in detail. It includes Overall Architecture(3.1),RMT-guided multi-scale fusion module(3.2), network layer deepening(3.3), multi-color space encoder(3.4), channel attention module(3.5), residual enhancement module(3.6) and loss function(3.7). 3.1.Network structure We have designed a multi-color space encoder network based on multi-scale fusion(MFUnet), and Fig. 2 shows the general architecture of the network. In this network, the input underwater image is first subjected to color space transformation, so that the input image enters three color coding paths: RGB path, Lab path and HSV path. In each color path, the input image is delivered to four serial enhancement modules. Each serial enhancement module performs a 2-fold down-sampling operation on the input to obtain four levels of feature representation. The three-color paths are tightly fused with the corresponding features of the RGB path, the HSV path and the Lab path while the serial enhancement is performed. The RGB paths are enhanced by replacing the original features of the RGB channels with the fused features. Subsequently, three parallel identical channel level features are concatenated, resulting in four sets of multi-color space encoder features, represented as [C2,C3,C4,C5]. And these four sets of color features are fed into the channel attention module of the corresponding inlet, which serves to extract the most representative semantic features. After obtaining the [C2,C3,C4,C5], they are input into the multi-scale fusion module guided by the medium transfer map (RMT). As shown in Fig. 3 , the extraction and fusion between multilevel features is achieved. The specific way to implement this is that the bottom features C2 and C3 are first fed to the feature pyramid. Then the features are weighted using an RMT map of the same size to compensate for quality degradation regions, denoted as [C2`, C3`]. Then feature fusion is performed on [C2`, C3`] to obtain a set of outputs. Then C4 is added and finally C5 is added to get the final multi-scale features weighted by the RMT map, represented as [P2, P3, P4, P5]. In the feature pyramid, RMT maps of different scales are implemented by maximum pooling. Ordinary convolution kernel is used for feature up-sampling and bilinear interpolation is used for feature down-sampling. Finally, the [P2,P3,P4,P5] of the output of the multi-scale fusion module are fed to the corresponding residual enhancement modules. After four serial residual modules with three times 2-fold up-sampling, the resulting decoded features are fed to the convolutional layer for the resultant reconstruction, resulting in an enhanced image. 3.2.RMT-guided asymptotic feature pyramid network Figure 3 shows our proposed RMT-guided asymptotic feature pyramid network (R-AFPN) process. To enhance the clarity and richness of details and textures in the resultant underwater images, while suppressing noise and artifacts, and improving regions of image degradation, we utilize the Reverse-Asymptotic Feature Pyramid Network (R-AFPN) module for feature fusion across different scales. Given that lower-layer features possess higher resolution but smaller receptive fields, and upper-layer features have larger receptive fields but lower resolution, we employ a hierarchical, multi-layer fusion strategy. This approach enriches the receptive fields of all-layer features, allowing each layer to assimilate semantic information from others, thereby bridging the semantic gap and boosting the overall network performance. However, a challenge arises during multi-scale fusion: the degraded regions within underwater images tend to merge, potentially exacerbating image degradation. To address this, we apply Reverse Medium Transmission (RMT) maps at varying scales to pre-compensate for feature degradation prior to fusion at each layer. This solution effectively slove the issue, enhancing the network's sensitivity to degraded areas and overall performance. Using RMT for weighted compensation is shown in Fig. 5 . Since the semantic gap between non-adjacent hierarchical features is larger than the semantic gap between adjacent hierarchical features, especially the gap between the bottom and the top, it directly leads to poor fusion of non-adjacent hierarchical features. Therefore it is not reasonable to use [C2,C3,C4,C5] directly for feature fusion. Therefore, we adopt progressive multi-scale fusion, which makes the semantic information of different level features closer together again in the process of progressive fusion, thereby alleviating the above problems. Our approach is to extract the last layer of each feature layer of multi-color channel encoding to obtain a set of feature [C2,C3,C4,C5] of different scales. In the bottom-up feature extraction process of the backbone network, the R-AFPN module gradually performs RMT degradation compensation on the bottom, high-level and top-level features and fuses them with each other. Specifically, the bottom features are first compensated for RMT degradation and fused with the bottom features to obtain new bottom features. Then the depth and the new bottom features are compensated for RMT degradation and fused, and the same is done for the top features. We use 1×1 convolution and bilinear interpolation methods to up-sample features, and use different convolution kernels and steps to perform down-sampling. After each feature fusion is completed, a residual block is used for feature learning for each feature layer. Each residual block consists of two 3×3 convolutions. For the specific fusion process, we assign different spatial weights to the features at different levels in the multilevel feature fusion process. In this way, we enhance the importance of key levels and reduce the influence of interference information to achieve the purpose of self-adaptation. The specific details are shown in Fig. 4 . The image formation model 44,45 under severe weather is widely used in image dehazing and underwater image restoration algorithms 10,32,35 , and its quality degradation image can be expressed as: $${\text{I}}^{\text{c}}\left(\text{x}\right)={\text{J}}^{\text{c}}\left(\text{x}\right)\otimes \text{T}\left(\text{x}\right)\oplus {\text{A}}^{\text{c}}\left(\text{x}\right)\otimes \left(1-\text{T}\left(\text{x}\right)\right),\text{c}\in \{\text{r},\text{g},\text{b}\},$$ 1 In the formula, x represents the pixel index, I is the observation image, J is the clear image, A is the uniform background light, and T is the medium projection rate, which means the percentage of scene brightness that reaches the camera after being reflected in the medium, reflecting the quality degradation in different areas degree. We incorporate the medium transmission map into the multi-scale fusion network via the proposed medium transmission guidance module. Specifically, the reverse medium transport (RMT)map,represented as \(\overline{\text{T}}\in {\mathbb{ℝ}}^{\text{H}\times \text{W}}\) ,is used as a pixel-level attention map. \(\overline{\text{T}}\) in the RMT diagram is obtained from 1- \(\text{T}\) \((\text{T}\in {\mathbb{ℝ}}^{\text{H}\times \text{W}}.\text{I}\text{t} \text{i}\text{s} \text{r}\text{e}\text{p}\text{r}\text{e}\text{s}\text{e}\text{n}\text{t}\text{s} \text{t}\text{h}\text{e} \text{T} \text{i}\text{s} [\text{0,1}]\) range media transfer map \(, \text{a}\text{n}\text{d} 1\in {\mathbb{ℝ}}^{\text{H}\times \text{W}}\) represents matrices with all elements 1) This means that more degraded areas should be allocated greater attention. However, real medium transport maps corresponding to underwater images are lacking in practice, so the depth model cannot be trained for medium transport map estimation. Inspired by robust general dark channel prior 35 , we estimate the medium transport map as: $$\stackrel{\sim}{\text{T}}\left(\text{x}\right)={\text{m}\text{a}\text{x}}_{\text{c},\text{y}\in {\Omega }\left(\text{x}\right)}\left(\frac{{\text{A}}^{\text{c}}-{\text{I}}^{\text{c}}\left(\text{y}\right)}{\text{m}\text{a}\text{x}\left({\text{A}}^{\text{c}},1-{\text{A}}^{\text{c}}\right)}\right),$$ 2 where \(\stackrel{\sim}{\text{T}}\) is the estimated medium transmission map, \({\Omega }\left(\text{x}\right)\) represents a local patch with a size of 15 × 15 centered on x, and c denotes the color channe, and A represents the uniform background light. In the method 35 , the estimation of uniform background light is based on depth-dependent color changes. Based on the RMT map, the principle of the proposed medium transport bootstrap module is shown in Fig. 5 . As can be seen from Eq. We use the RMT map as a feature selector to weight the different spatial locations of the features. Pixels with high degradation (pixels with larger RMT values) will be given higher weights, which can be expressed as: $$\mathcal{V}=\mathcal{U}\oplus \mathcal{U}\otimes \overline{\text{T}},$$ 3 The \(\mathcal{V}\in {\mathbb{ℝ}}^{\text{M}\times \text{H}\times \text{W}} \text{a}\text{n}\text{d} \mathcal{U}\in {\mathbb{ℝ}}^{\text{M}\times \text{H}\times \text{W}}\) in the formula represent the output features and input features of the media transmission guidance module respectively. We treat the RMT weight as an identity connection to avoid vanishing gradients and tolerate errors caused by inaccurate estimation of medium transmission. Furthermore, our purely data-driven framework is also tolerant of inaccuracies in the media transfer graph. Inspired by the method 24 , we use an adaptive spatial fusion algorithm to enhance the importance of key levels and mitigate the effect of contradictory information from different objects. As shown in Fig. 4 , we fused three levels of features. \({\text{x}}_{\text{i}\text{j}}^{\text{n}\to \text{l}}\) represents the feature vectors at the level from \(\text{n}\) level to \(\text{l} \text{l}\text{e}\text{v}\text{e}\text{l}\left(\text{i},\text{j}\right)\) . The \({\text{y}}_{\text{i}\text{j}}^{\text{l}},\) feature vector obtained by adaptive spatial fusion of multilevel features is obtained by \({\text{x}}_{\text{i}\text{j}}^{1\to \text{l}},{\text{x}}_{\text{i}\text{j}}^{2\to \text{l}},{\text{x}}_{\text{i}\text{j}}^{3\to \text{l}}\) linear combination. It is shown below: $${\text{y}}_{\text{i}\text{j}}^{\text{l}}={{\alpha }}_{\text{i}\text{j}}^{\text{l}}\cdot {\text{x}}_{\text{i}\text{j}}^{1\to \text{l}}+{{\beta }}_{\text{i}\text{j}}^{\text{l}}\cdot {\text{x}}_{\text{i}\text{j}}^{2\to \text{l}}+{{\gamma }}_{\text{i}\text{j}}^{\text{l}}\cdot {\text{x}}_{\text{i}\text{j}}^{3\to \text{l}},$$ 4 In the formula, \({{\alpha }}_{\text{i}\text{j}}^{\text{l}},{{\beta }}_{\text{i}\text{j}}^{\text{l}}\text{与}{{\gamma }}_{\text{i}\text{j}}^{\text{l}}\) represents the spatial weight of the three-level features of level \(\text{l}\) , subject to the \({{\alpha }}_{\text{i}\text{j}}^{\text{l}}+{{\beta }}_{\text{i}\text{j}}^{\text{l}}+{{\gamma }}_{\text{i}\text{j}}^{\text{l}}=1\) constraint. 3.3.Deepening of the network hierarchy The structure of U-net 23 network facilitates the integration of information from the various stages of down-sampling in the up-sampling process, so that the high level of the network obtains the high-frequency information of the graph, and the bottom level of the network obtains the low-frequency information of the graph. That is, the up-sampling process combines the structural information of the various levels. The information at each level is then preserved using jump connections. The traditional network structure 23 is usually a three-layer network structure, but after our experiments, we learned that deepening the network layers can further play the advantages of multi-color encoder to enhance the color spatial representation of the image, and play the advantages of multi-scale fusion to reduce the semantic gap of each level through increasing the sensitivity of the restored image to the degraded region, and improve the model's ability to generalize and enhance the performance of the image. The results are shown in the ablation experiment. 3.4.Multi-color space encoder According to the underwater imaging model, images captured underwater exhibit significant color deviations, often appearing bluer, greener, or yellower than those observed on land. The diverse color spectrum of underwater images often renders traditional network architectures ineffective. To address this challenge, it is essential to adopt a novel network architecture that ensures enhanced underwater images appear more natural. Inspired by some multi-color enhancement algorithms 22 working in color spaces, an image will have different expressions in different color spaces, so we extracted the features from three color spaces of the image, RGB, HSV and Lab, in order to correct the color deviation of the underwater image. Currently, the RGB color space is the most prevalent, primarily because the human eye predominantly perceives the three colors of red, green, and blue, making operations within the RGB color space relatively straightforward and amenable to mathematical processing. However, the components of R, G, and B are highly interrelated and significantly influenced by variations in brightness and shadow, complicating color correction through other methods.In the HSV color space, the hue, saturation, brightness and contrast of the image can be intuitively reflected. In the Lab color space, all colors that can be observed by the human eye can be displayed. The characteristics of each color space are very obvious, and the correlation between color spaces is also low, which provides prerequisites for color correction of underwater images. We integrate features from various color spaces into a cohesive depth structure, enabling the utilization of components such as color, hue, and saturation, which are pertinent to image degradation. Importantly, the color disparity between two adjacent points might be negligible within one color space, yet pronounced in others. Therefore, the embedding of multi-color space makes it easier to measure the color deviation of an underwater image and correct the image. In addition, the multi-color encoder brings more non-linear operations in the color space transformation process, which significantly improves the performance of the depth model. 3.5.Channel attention module Given the distinct definitions of each color space, it is crucial to assign appropriate weights to the features extracted from the three color spaces to optimize the functionality of the multi-color encoder. To achieve this, we utilize a channel attention module to illustrate the interdependence among different color spaces. The channel attention module is shown in Fig. 6 . Suppose the input feature is \(\begin{array}{r}\mathcal{ℱ}=\mathsf{C}\mathsf{a}\mathsf{t}\left({\text{F}}_{1},{\text{F}}_{2},\cdots ,{\text{F}}_{\text{N}}\right)\in \end{array}{ \mathbb{ℝ}}^{\text{N}\times \text{H}\times \text{W}}\) ,F is the feature mapping of a path at a particular level; N is the number of feature mappings; Cat is the feature stitching; H and W are the height and width of the input image, respectively. We first perform global average pooling on the input features F to obtain the channel descriptor \(\text{z}\in {\mathbb{ℝ}}^{\text{N}\times 1}\) ,followed by the embedded global distribution of the channel feature responses.The k-th item of z can be expressed as: $$\begin{array}{r}{\text{z}}_{\text{k}}=\frac{1}{\text{H}\times \text{W}}\sum _{\text{i}}^{\text{H}}\sum _{\text{j}}^{\text{W}}{\text{F}}_{\text{k}}\left(\text{i},\text{j}\right),\end{array}$$ 5 In the formula \(\text{k}\in \left[1,\text{N}\right]\) . In order to fully extract the dependencies of each channel, we used self-gating mechanism 46 to obtain the set of modulation weights for each channel \(\text{s}\in {\mathbb{ℝ}}^{\text{N}\times 1}\) : $$\text{s}={\sigma }\left({\text{W}}_{2}\ast \left({\delta }\left({\text{W}}_{1}\ast \text{z}\right)\right)\right),$$ 6 Where \({\sigma }\left(\cdot \right)\) represents the Sigmoid activation function, \({\delta }\left(\cdot \right)\) represents the ReLU activation function, * represents the convolution operation, W1 and W2 are the weights of the two fully connected layers respectively, and their output channel numbers are \(\frac{\text{N}}{\text{r}}\) and N,respectively, where r is 16, in order to reduce the computational cost. Finally, these weights will be applied to the input features F to generate the rescaled features \(\mathcal{U}\in {\mathbb{ℝ}}^{\text{N}}\times \text{H}\times \text{W}\) .Furthermore, in order to maintain the good properties of the original features while avoiding the problem of gradient vanishing, we treat the channel-attention weights in the same mapping manner: $$\mathcal{U}=\mathcal{ℱ}\oplus \mathcal{ℱ}\otimes \text{s},$$ 7 where \(\oplus\) and \(\otimes\) denote pixel-by-pixel addition and pixel-by-pixel multiplication, respectively. 3.6.Residual enhancement module Figure 7 shows the residual enhancement module we used. The primary objective of the residual enhancement module is to ensure the integrity and fidelity of the data, while also preventing the occurrence of gradient explosion or vanishing. To enhance the overall performance and stability of the network, we strategically place residual enhancement modules throughout the network, at each up-sampling and down-sampling layer. Within these modules, the convolutional layers are standardized to have an identical number of filters. In each residual enhancement module, the convolutional layers have the same number of filters. In the encoder network, the number of filters is gradually increased from 16 to 128 by a factor of 2. In the decoder network, the number of filters is reduced from 128 to 16 by a factor of 2. All the convolutional layers have a convolutional kernel of 3 × 3 and a step size of 1. 3.7.Loss function We use linear combination of \({\text{l}}_{2}\) loss \({L}_{{l}_{2}}\) , perceptual loss \({\text{L}}_{\text{p}\text{e}\text{r}}\) and edge loss \({\text{L}}_{\text{e}\text{d}\text{g}}\) to balance visual quality and quantitative scores: $$L={\lambda }_{1}{L}_{{l}_{2}}+{\lambda }_{2}{L}_{per}+{\lambda }_{3}{L}_{edg}$$ 8 Where \({L}_{{l}_{2}}\) represents the loss \({l}_{2}\) ; \({L}_{per}\) represents the perceptual loss; \({L}_{edg}\) represents the edge loss; \({\lambda }_{1}\) , \({\lambda }_{2}\) and \({\lambda }_{3}\) are the balancing factors. In this chapter let \({\lambda }_{1}\) =5, \({\lambda }_{3}\) =0.05, \({\lambda }_{3}\) =10. \({l}_{2}\) loss \({L}_{{l}_{2}}\) measures the difference between the reference image \(I\) and the reconstructed image : $${L}_{{l}_{2}}=\sum _{m=1}^{H}\sum _{n=1}^{W}{\left(\widehat{I}\left(m,n\right)-I\left(m,n\right)\right)}^{2}$$ 9 In the formula, \(I\) and \(\widehat{I}\) are the reference and reconstructed images, respectively; \(H\) and \(W\) are the height and width of the images. The perceptual loss \({L}_{per}\) combines the features extracted from the VGG-19 47 network to measure the structural consistency of the reconstructed image and the reference image. Perceptual loss is defined as follows: $${L}_{per}=\sum _{m=1}^{H}\sum _{n=1}^{W}\left|{\varphi }_{i}\left(\widehat{I}\right)\left(m,n\right)-{\varphi }_{i}\left(I\right)\left(m,n\right)\right|$$ 10 In the formula, \(\varphi\) represents the VGG-19 network pre-trained on the ImagNet dataset, and \({\varphi }_{i}\left(\cdot \right)\) is the i-th convolutional layer. Here we use the features output by the relu5_4 layer of the VGG-19 network. The edge loss \({L}_{edg}\) is used to reconstruct the edges and texture of the image and the edge loss is denoted as: $${L}_{edg}=\sqrt{\parallel \varDelta \widehat{I}-\varDelta I{\parallel }^{2}+{\epsilon }^{2}}$$ 11 Where \(\varDelta\) denotes the Laplace operator and \(\epsilon\) is set to \({10}^{-3}\) . 4. Experiments Next, the implementation details of this chapter are described first, and then the software and hardware environment and experimental settings of the experiment are introduced. We compared the method proposed in this chapter with representative methods and performed ablation experiments to verify the effectiveness of each part. 4.1.Data sets To be fair, our training set and benchmarks are consistent with the literature 41 . They are introduced separately below. ( 1 )Training set. To train the algorithms proposed in this chapter, we randomly selected 800 pairs of underwater images from the UIEB 41 underwater image enhancement dataset. The UIEB dataset consists of 890 pairs of degraded underwater pictures and the corresponding reference versions, which cover different underwater scenes and different degradation situations. We also selected 1250 images from a synthetic underwater dataset 21 , divided into subsets of 10 water quality types, including open seawater as I, IA, IB, II, III, and coastal waters 1, 3, 5, 7, and 9. Additionally, to increase the training data, we randomly cropped the patches with a size of 128x128. ( 2 )Benchmarks. We use the remaining 90 pairs of images from the UIEB dataset, denoted as Test-R90. we select 100 pairs of images from each subset of the synthetic underwater dataset 21 , denoted as Test-R1000. In addition, we also conduct comprehensive experiments on Test-C60 41 , SQUID 48 and Color-Check7 27 . Test-C60 contains 60 real underwater images that do not have the reference images provided in the UIEB. Test-C60 is more challenging compared to Test-R90. The Squud dataset contains 57 pairs of underwater images from 4 different dive sites in Israel. We used 16 representative examples provided on the SQUID project page for testing. Specifically, 4 representative samples from each of the 4 dive sites (Katzaa, Michmoret, Nachsholim, Satil) were selected, and the resolution of each image was 1827×2737.Color-Check7 contains 7 underwater color checker images taken with different cameras provided in the literature, and was used to evaluate the underwater robustness and accuracy of color correction. The cameras used to take the color-check photographs are denoted in this paper as Can D10, Fuj Z33, Oly T6000, Oly T8000, Pen W60, Pen W80, and Pan TS1. ( 3 )Comparison of methods. We compare with 11 other typical underwater image enhancement methods. These include a non-physical modelling approach (Ancutiet al 26 .), three physical model-based approaches, (Li et al 12 ., Peng et al 32 ., GDCP 35 ), four learned approaches, (UcycleGAN 19 , Guo et al 49 ., Water-Net 20 ,UWCNN 21 ) and three baseline depth models(Unet-U 23 、Unet-RMT 23 and Ucolor 22 ). Due to hardware platform limitations, both the methods proposed in this chapter and Ucolor training reduce the batch size to 10. This undoubtedly reduces the actual measured performance data, but does not affect the conclusions of the final experiments. The data for the other compared methods are taken from the publicly available data in literature 41 . 4.2.Evaluation metrics. For Test-R90 and Test-S1000, we used PSNR, SSIM and MSE for evaluation. Similar to the literature 41 , the reference image of Test-R90 is used as the real image in order to calculate the PSNR, SSIM and MSE scores. A higher PSNR score or a higher SSIM score or a lower MSE score indicates that the resultant image is closer to the image content of the real image and the algorithm performs better. For Test-C60 and SQUID that do not have corresponding ground truth images, we employ the no-reference evaluation metrics UCIQE 50 and UIQM 51 to measure the performance of different methods. The Higher UCIQE or UIQM scores indicate that the image is visually better. In addition, we also provide NIQE 52 test scores. The lower NIQE scores indicate better image quality. Due to hardware limitations, we scaled the images in the SQUID test set to 512×512 and then processed them with different algorithms. For Color-Check7, we tested the difference in swatch colors between the real image and the corresponding enhanced image. Specifically, we cropped 24 color blocks from the enhancement result and the corresponding real image and calculated the average color value of each block separately. Then the method in literature 27 is adopted, and CIEDE 53 is used to measure the relative perceptual difference between corresponding color patches. The smaller the value of CIEDE, the better the image enhancement effect. 4.3.Implementation details The software environment we used is Python 3.6, TensorFlow 1.9, cudnn 7.1.4, cuda10.0. The hardware platform is Intel Xeon E5-2680 for CPU, 32GB of RAM, and Titan Xp for GPU with 12GB of video memory. The batch size for training is 10, and the training period is 500. The channel base is 16, initialized with Gaussian distribution, the optimizer is Adam, the learning rate is fixed to 1e-4 and the training time is one day. 4.4.Visual comparisons In this section, we conduct a visual comparison across different test sets, beginning with an evaluation of the visualization effects on synthetic images. As depicted in Fig. 8 , the algorithm introduced in this chapter outperforms the competing algorithms in terms of image dehazing and color correction. Among all the algorithms assessed, the one proposed in this study most closely approximates the scene structure of the actual image, achieving the highest Peak Signal-to-Noise Ratio (PSNR) ,and the lowest Mean Squared Error (MSE) values. In Fig. 9 , we show the results of a real underwater image with a significant green color bias sampled from Test-R90 processed by different algorithms. The color bias in the degraded underwater image obscures the details of the image, and correcting the color bias results in clearer details. In terms of color correction, all the comparison methods are unsatisfactory, with the colors showing either under or over compensation. In particular, GDCP and Ancuti introduce additional color artefacts. In contrast, the method proposed in this chapter effectively eliminates color bias, improves contrast, and without obvious under-compensation, over-compensation or artifacts. We also sampled challenging images from Test-C60 for comparison, which generally suffer from severe degradation such as color shift, blurring, low brightness, etc., and the results of different methods of enhancement are shown in Fig. 10 . In all the results showcased, none of the compared methods achieved satisfactory outcomes; some algorithms produced artifacts, while others introduced artificial colors. The highest Underwater Image Quality Measure (UIQM) was achieved by the algorithm proposed in this study. In Fig. 11 , we show the visual effect of sampled images in SQUID. Ancuti et al.'s algorithm achieved the best contrast, but the image showed a green color cast. The method proposed in this chapter can effectively improve the contrast of the image without producing obvious artificial color shift. We have compared the processing results of underwater Color Checker images and the results are shown in Fig. 12 . All the compared methods try to correct the color deviation as a whole, but the proposed method in this paper has the best color correction results. The comparisons of visual effects presented above demonstrate that the algorithm introduced in this study outperforms all other algorithms in terms of visual quality. Furthermore, it exhibits robustness across various datasets and shows excellent generalizability to different underwater scenes. 4.5.Quantitative comparisons In order to further verify the performance of the algorithm proposed in this chapter, we first conducted a quantitative comparison on the Test-S1000 and Test-R90 data sets. Due to hardware platform limitations, we change the training batch size of Ucolor to 10. The average PSNR, MSE (×1000) and SSIM scores of different methods are shown in Table 1 , and the best results for each indicator are marked in red. As shown in Table 1 , the performance of the algorithm proposed in this chapter is better than almost all compared methods. Compared with Ucolor, which ranks second in overall performance, the PSNR, MSE and SSIM scores on Test-S1000 are increased by 7.3%, 44.5% and 44.5% respectively. 3.0%. The PSNR, MSE and SSIM scores on Test-R90 are improved by 1.3%, 5.5% and 3.9% respectively. Table 1 Evaluation results of different algorithms on TEST-S1000 and TEST-R90, "-" indicates that the results are not available, and “*” indicates that the training batch is 10. Methods Test-S1000 Test-R90 PSNR↑ MSE↓ SSIM↑ PSNR↑ MSE↓ SSIM↑ input 12.96 4.60 - 16.11 2.03 - Ancuti et al 26 . 13.27 5.15 0.51 19.19 0.78 0.79 Li et al 13 . 14.29 3.64 - 16.73 1.38 - Peng et al 32 . 13.04 4.53 0.18 15.77 1.72 0.56 GDCP 35 11.67 5.98 0.26 13.85 3.40 0.50 Guo et al 49 . 15.78 2.57 - 18.05 1.18 - UcycleGAN 19 14.73 3.13 - 16.61 1.65 - Water-Net 41 15.47 3.26 0.53 19.81 1.02 0.75 UWCNN type1 21 16.27 2.68 - 13.62 3.52 - UWCNN type3 21 15.70 2.87 - 12.84 4.23 - UWCNN type5 21 14.78 2.94 - 13.26 3.65 - UWCNN type7 21 12.38 4.35 - 13.02 3.67 - UWCNN type9 21 12.83 3.85 - 12.79 3.89 - UWCNN typeI 21 10.44 6.42 - 10.57 6.24 - UWCNN typeII 21 17.51 2.59 - 14.75 2.57 - UWCNN typeIII 21 17.41 2.39 - 13.26 3.40 - UWCNN retrain21 15.87 2.74 - 16.69 1.71 - Unet-U 23 19.14 1.22 - 18.14 1.32 - Unet-RMT 23 17.93 1.43 - 16.89 1.71 - Ucolor* 20.19 1.19 0.76 20.18 0.91 0.74 Ours 21.66 0.66 0.78 20.44 0.86 0.77 We tested the performance of different algorithms on the Test-C60 and SQUID datasets. Table 2 shows the average scores of UIQM, UCIOE and NIQE for different algorithms. The best results for each are marked in red. It can be seen that none of the algorithms outperforms the others on these two challenging datasets. Overall, Li et al.'s algorithm performs the best, and the performance of the algorithms proposed in this chapter is in the middle of the pack overall. Table 2 UIQM, UCIQE and NIQE scores of different algorithms on Test-C60 and SQUID datasets. "-" indicates that the result is not available and “*” indicates a training batch of 10. Methods Test-C60 SQUID UIQM↑ UCIQE↑ NIQE↓ UIQM↑ UCIQE↑ NIQE↓ input 0.84 0.48 7.14 0.82 0.42 4.93 Ancuti et al 26 . 1.22 0.62 4.94 1.3 0.62 5.01 Li et al 13 . 1.27 0.65 5.32 1.34 0.66 4.81 Peng et al 32 . 1.13 0.58 6.01 0.99 0.5 4.39 GDCP 35 1.07 0.56 5.92 1.11 0.52 4.48 Guo et al 49 . 1.11 0.6 5.71 - - - UcycleGAN 19 0.91 0.58 7.67 1.11 0.56 5.93 Water-Net 41 0.97 0.56 6.04 1.03 0.54 4.72 UWCNN typeII 21 0.77 0.47 6.76 0.69 0.44 4.60 UWCNN retrain 21 0.84 0.49 6.66 0.77 0.46 4.38 Unet-U 23 0.94 0.5 6.12 0.82 0.5 4.38 Unet-RMT 23 1.03 0.52 6.12 0.82 0.49 5.16 Ucolor* 0.94 0.53 7.49 0.77 0.50 7.50 Ours 1.096 0.54 6.85 0.94 0.52 7.26 Next, we compared the underwater check colour card images in Color-Check7 to demonstrate the accuracy and robustness of the colour correction of the method proposed in this chapter, and the results are shown in Table 3 . The best results for each are marked in red. The performance of the algorithm proposed in this chapter is optimal in most cases (6 cameras), and only slightly worse than its competitors on Pen W80 and Pan TS1. Table 3 CIEDE2000(↓) scores of different algorithms on Color-Check7, “*” indicates the training batch is 10 Methods Pen W60 Pen W80 Can D10 Fuj Z33 Oly T6000 Oly T8000 Pan TS1 Avg input 13.82 17.26 16.13 16.37 14.89 23.14 19.06 17.24 Ancuti et al 26 . 12.48 13.30 14.28 11.43 11.57 12.58 10.63 12.32 Li et al 13 . 15.41 17.56 18.52 25.01 16.01 17.12 12.03 17.38 Peng et al 32 . 13.16 16.01 14.78 14.09 12.24 14.79 19.59 14.95 GDCP 35 15.49 24.32 16.89 13.73 12.76 16.82 12.93 16.13 Guo et al 49 . 12.29 15.50 14.58 16.65 39.71 15.14 12.40 18.04 UcycleGAN 19 21.19 21.23 22.96 26.28 20.88 23.42 19.02 22.14 Water-Net 41 12.51 19.57 15.44 12.91 17.55 21.73 18.84 16.94 UWCNN typeII 21 16.73 20.55 17.73 17.20 16.31 17.94 20.97 18.20 UWCNN retrain 21 13.64 20.33 14.91 13.38 14.72 18.11 20.19 16.47 Unet-U 23 11.22 15.17 13.32 11.91 10.87 15.12 17.31 13.56 Unet-RMT 23 12.37 19.01 15.57 14.80 13.26 16.47 19.55 15.86 Ucolor* 10.54 9.60 12.14 9.20 11.37 13.94 12.06 11.26 Ours 9.56 10.02 10.60 7.88 8.35 11.63 13.83 10.27 4.6.Ablation study In order to analyze the effectiveness of each core component of the method proposed in this chapter, we conducted an ablation study on the Test-R90 data set. It includes RMT-guided multi-scale feature fusion (R-AFPN), 4-level refined network structure (4-layers) and edge loss (edge-loss). As shown in Table 4 , the complete model proposed in this chapter achieved the best results, demonstrating the effectiveness of the method proposed in this article. The results for each of the other core component combination methods also confirmed the necessity of each component. Table 4 results of ablation experiments on Test-R90 R-AFPN 4-layers edge-loss PSNR (↑) MSE×10³ (↓) SSIM (↑) UIQM (↑) UCIQE (↑) NIQE (↓) × × × 20.005 0.968 0.745 1.385 0.558 7.996 √ × × 20.356 0.832 0.742 1.422 0.564 7.742 × √ × 19.974 0.955 0.766 1.461 0.566 7.617 × × √ 19.965 0.976 0.745 1.441 0.561 7.197 √ √ × 20.413 0.862 0.773 1.441 0.567 7.834 √ × √ 20.037 0.929 0.767 1.404 0.563 7.331 × √ √ 19.971 0.917 0.745 1.337 0.554 7.416 √ √ √ 20.439 0.862 0.767 1.542 0.571 7.410 4.7.Comparison with Ucolor 22 To further exemplify our improvements, Table 5 shows a comparison of the model size and other performances of our proposed network with the Ucolor network. In order to make the network work on embedded platforms with limited hardware, we reduced the training parameters of the network, achieved by changing the channel base from 128 to 16, which reduces the enhancement effect but in return results in a lightweight network. In the comparison of model size, Ours is reduced by 73.05% compared to the Ucolor network. In terms of training time, Ours takes only 1/3 of the time required by the Ucolor network, and is 8.89% faster than the Ucolor network in terms of average speed of image processing. Table 5 Comparison with the Ucolor network, “*” indicates the training batch is 10*, “C60”indicates testing on the Test-C60 dataset Compare items Ucolor* Ours Model size 756.09MB 203.78MB Training duration Three days One days Processing speed C60 171.0s 155.8s Channel base 128 16 5. Conclusions We introduce an underwater image enhancement method leveraging multi-scale feature fusion. Traditional approaches fail to address the issue of information loss and degradation due to the indirect interaction between non-adjacent layers, nor do they fully harness the potential of each module. Moreover, most existing image enhancement techniques focus on spatial domain enhancement, neglecting frequency domain information enhancement. Our method employs a progressive feature pyramid network to overcome these challenges. Specifically, we adopt progressive strategies and self-adaptation for feature fusion. During the fusion process, the medium transmission map is utilized as a weight of importance, integrating underwater imaging domain knowledge into the network. We enhance the network's coding and decoding depth to maximize the benefits of each module and employ edge loss as a penalty term to amplify frequency domain detail enhancement. Extensive testing on various benchmarks confirms the superiority of our approach. However, our method exhibits certain limitations. Firstly, the enhancement of frequency domain details is primarily focused on edge loss, resulting in suboptimal performance in enhancing frequency domain details. Secondly, the estimation of the medium transmission map, which guides the multi-scale fusion, is occasionally inaccurate, thereby impacting the overall enhancement effect. To address these issues, future efforts will concentrate on augmenting the frequency domain enhancement techniques and refining the guidance method to further improve our approach. Declarations Data availability The UIEB dataset is available at https://li-chongyi.github.io/proj_benchmark.html with credentialed access. Acknowledgements This work was supported by the College Student Innovation and Entrepreneurship Project of Hainan University(Hdcxcyxm201704), the Hainan Provincial Natural Science Foundation of China (623RC449). Author contributions Y.H: Conceptualization, supervision, methodology, software, validation, formal analysis, writing-original draft, writing-review&edition; C.H.Q.: conceptualization, methodology, data curation, software, validation, formal analysis, writing-original draft, writing-review&edition; J.C.X.: data curation, resources, writing-review&edition, visualization; Z.R.T.: resources, funding acquisition; Z.J.: supervision, writing-review&edition, project administration, funding acquisition; All authors read and approved the final article. Competing interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. References Gao, S. B., Zhang, M., Zhao, Q., Zhang, X. S. & Li, Y. J. Underwater Image Enhancement Using Adaptive Retinal Mechanisms. IEEE Trans Image Process 28, 5580–5595, doi: 10.1109/TIP.2019.2919947 (2019). Singhai, J. & Rawat, P. in International Conference on Computational Intelligence and Multimedia Applications (ICCIMA 2007). 507–512 (IEEE). Zhang, W., Dong, L., Zhang, T. & Xu, W. Enhancing underwater image via color correction and bi-interval contrast enhancement. Signal Processing: Image Communication 90, 116030, doi: 10.1109/vcip.2017.8305027 . (2021). Zhuang, P. & Ding, X. Underwater image enhancement using an edge-preserving filtering retinex algorithm. Multimedia Tools Applications 79, 17257–17277, doi: 10.1007/s11042-020-08739-3 . (2020). Drews, P., Nascimento, E., Moraes, F., Botelho, S. & Campos, M. in Proceedings of the IEEE international conference on computer vision workshops. 825–830. Song, W., Wang, Y., Huang, D. & Tjondronegoro, D. in Advances in Multimedia Information Processing–PCM 2018: 19th Pacific-Rim Conference on Multimedia, Hefei, China, September 21–22 , 2018, Proceedings, Part I 19. 678–688 (Springer). Song, W., Wang, Y., Huang, D., Liotta, A. & Perra, C. Enhancement of underwater images with statistical model of background light and optimization of transmission map. IEEE Transactions on Broadcasting 66, 153–169, doi: 10.1109/TBC.2019.2960942 . (2020). Yu, H. et al. Underwater image enhancement based on DCP and depth transmission map. Multimedia Tools and Applications 79, 20373–20390, doi: 10.1007/s11042-020-08701-3 . (2020). Zhuang, P., Li, C. & Wu, J. Bayesian retinex underwater image enhancement. Engineering Applications of Artificial Intelligence 101, 104171, doi: 10.1016/j.engappai.2021.104171 . (2021). Chiang, J. Y. & Chen, Y.-C. Underwater image enhancement by wavelength compensation and dehazing. IEEE transactions on image processing 21, 1756–1769, doi: 10.1109/TIP.2011.2179666 . (2012). Drews, P. L., Nascimento, E. R., Botelho, S. S. & Campos, M. F. M. Underwater depth estimation and image restoration based on single images. IEEE computer graphics & applications 36, 24–35, doi: 10.1109/MCG.2016.26 . (2016). Li, C.-Y., Guo, J.-C., Cong, R.-M., Pang, Y.-W. & Wang, B. Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior. IEEE Transactions on Image Processing 25, 5664–5677, doi: 10.1109/TIP.2016.2612882 . (2016). Li, C. et al. in IEEE International Conference on Image Processing (ICIP). 1993–1997 (IEEE). Berman, D., Treibitz, T. & Avidan, S. in Proc. British Machine Vision Conference (BMVC). 2. Li, C., Guo, J., Guo, C., Cong, R. & Gong, J. A hybrid method for underwater image correction. Pattern Recognition Letters 94, 62–67, doi: 10.1016/j.patrec.2017.05.023 . (2017). He, K., Sun, J. & Tang, X. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis 33, 2341–2353, doi: 10.1109/TPAMI.2010.168 . (2011). Galdran, A., Pardo, D., Picón, A. & Alvarez-Gila, A. Automatic red-channel underwater image restoration. Journal of Visual Communication and Image Representation 26, 132–145, doi:10.1016/j.jvcir.2014.11.006. (2015). Drews, P. L., Nascimento, E. R., Botelho, S. S., Campos, M. F. M. J. I. c. g. & applications. Underwater depth estimation and image restoration based on single images. 36, 24–35, doi: 10.1109/MCG.2016.26 (2016). Li, C., Guo, J. & Guo, C. Emerging from water: Underwater image color correction based on weakly supervised color transfer. IEEE Signal processing letters 25, 323–327, doi: 10.1109/LSP.2018.2792050 . (2018). Li, J., Skinner, K. A., Eustice, R. M. & Johnson-Roberson, M. WaterGAN: Unsupervised generative network to enable real-time color correction of monocular underwater images. IEEE Robotics and Automation letters 3, 387–394, doi: 10.1109/LRA.2017.2730363 . (2018). Li, C., Anwar, S. & Porikli, F. Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition 98, 107038, doi: 10.1016/j.patcog.2019.107038 (2020). Li, C. et al. Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE Transactions on Image Processing 30, 4985–5000, doi: 10.1109/TIP.2021.3076367 (2021). Ronneberger, O., Fischer, P. & Brox, T. in Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5–9 , 2015, Proceedings, Part III 18. 234–241 (Springer). Yang, G. et al. in IEEE International Conference on Systems, Man, and Cybernetics (SMC). 2184–2189 (IEEE). Huo, F., Li, B. & Zhu, X. in Proceedings of the IEEE/CVF International Conference on Computer Vision. 1944–1952. Ancuti, C., Ancuti, C. O., Haber, T. & Bekaert, P. in IEEE conference on computer vision and pattern recognition. 81–88 (IEEE). Ancuti, C. O., Ancuti, C., De Vleeschouwer, C. & Bekaert, P. Color balance and fusion for underwater image enhancement. IEEE Transactions on image processing 27, 379–393, doi: 10.1109/TIP.2017.2759252 (2017). Ancuti, C. O., Ancuti, C., Haber, T. & Bekaert, P. in IEEE International Conference on Image Processing. 1557–1560 (IEEE). Fu, X. et al. in 2014 IEEE international conference on image processing (ICIP). 4572–4576 (IEEE). Ghani, A. S. A. & Isa, N. A. M. Underwater image quality enhancement through integrated color model with Rayleigh distribution. Applied soft computing 27, 219–230, doi: 10.1016/j.asoc.2014.11.020 (2015). Drews, P. L., Nascimento, E. R., Botelho, S. S. & Campos, M. F. M. Underwater depth estimation and image restoration based on single images. IEEE computer graphics applications 36, 24–35, doi: 10.1109/MCG.2016.26 (2016). Peng, Y.-T. & Cosman, P. C. Underwater image restoration based on image blurriness and light absorption. IEEE transactions on image processing 26, 1579–1594, doi: 10.1109/TIP.2017.2663846 . (2017). Akkaynak, D. & Treibitz, T. in Proceedings of the IEEE conference on computer vision and pattern recognition. 6723–6732. Akkaynak, D. & Treibitz, T. in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1682–1691. Peng, Y.-T., Cao, K. & Cosman, P. C. Generalization of the dark channel prior for single image restoration. IEEE Transactions on Image Processing 27, 2856–2868, doi: 10.1109/TIP.2018.2813092 . (2018). Akkaynak, D. et al. in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 4931–4940. Fabbri, C., Islam, M. J. & Sattar, J. in 2018 IEEE international conference on robotics and automation (ICRA). 7159–7165 (IEEE). Uplavikar, P. M., Wu, Z. & Wang, Z. in CVPR workshops. 1–8. Jamadandi, A. & Mudenagudi, U. in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 11–17. Islam, M. J., Xia, Y., Sattar, J. J. I. R. & Letters, A. Fast underwater image enhancement for improved visual perception. 5, 3227–3234, doi: 10.1109/LRA.2020.2974710 (2020). Li, C. et al. An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing 29, 4376–4389, doi: 10.1109/TIP.2019.2955241 (2020). Jamadandi, A. & Mudenagudi, U. in CVPR Workshops. Jerlov, N. G. Marine optics . 1, 2 (Elsevier, 1976). Narasimhan, S. G. & Nayar, S. K. Vision and the atmosphere. International journal of computer vision 48, 233–254, doi: 10.1145/1508044.1508113 . (2002). Tan, R. T. in IEEE conference on computer vision and pattern recognition. 1–8 (IEEE). Hu, J., Shen, L. & Sun, G. in Proceedings of the IEEE conference on computer vision and pattern recognition. 7132–7141. Simonyan, K. & Zisserman, A. Very deep convolutional networks for large-scale image recognition. Computer Science, doi: 10.48550/arXiv.1409.1556 (2014). Berman, D., Levy, D., Avidan, S. & Treibitz, T. Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE transactions on pattern analysis and machine intelligence 43, 2822–2837, doi: 10.1109/TPAMI.2020.2977624 (2020). Guo, Y., Li, H. & Zhuang, P. Underwater image enhancement using a multiscale dense generative adversarial network. IEEE Journal of Oceanic Engineering 45, 862–870, doi: 10.1109/JOE.2019.2911447 (2019). Yang, M. & Sowmya, A. An underwater color image quality evaluation metric. IEEE Transactions on Image Processing 24, 6062–6071, doi: 10.1109/TIP.2015.2491020 (2015). Panetta, K., Gao, C. & Agaian, S. Human-visual-system-inspired underwater image quality measures. IEEE Journal of Oceanic Engineering 41, 541–551, doi: 10.1109/JOE.2015.2469915 (2015). Mittal, A., Soundararajan, R. & Bovik, A. C. Making a “completely blind” image quality analyzer. IEEE Signal processing letters 20, 209–212, doi: 10.1109/LSP.2012.2227726 (2012). Sharma, G., Wu, W. & Dalal, E. N. The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research & Application 30, 21–30, doi: 10.1002/col.20070 (2005). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4082073","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":284006092,"identity":"89d460e1-db5f-4519-b330-b51d9609c055","order_by":0,"name":"Hao Yang","email":"","orcid":"","institution":"Hainan University","correspondingAuthor":false,"prefix":"","firstName":"Hao","middleName":"","lastName":"Yang","suffix":""},{"id":284006094,"identity":"ddeffd08-506f-40a9-8399-a99259cdc5fa","order_by":1,"name":"Hongqin Cai","email":"","orcid":"","institution":"Hainan University","correspondingAuthor":false,"prefix":"","firstName":"Hongqin","middleName":"","lastName":"Cai","suffix":""},{"id":284006096,"identity":"d46feaed-79d8-44b7-be32-42fba7e2015d","order_by":2,"name":"Chenxu Jiang","email":"","orcid":"","institution":"Hainan University","correspondingAuthor":false,"prefix":"","firstName":"Chenxu","middleName":"","lastName":"Jiang","suffix":""},{"id":284006097,"identity":"d53ffd1a-6330-4aa2-8ae6-c933b665bbe2","order_by":3,"name":"Ruiteng Zhang","email":"","orcid":"","institution":"Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Ruiteng","middleName":"","lastName":"Zhang","suffix":""},{"id":284006098,"identity":"51d3ad50-0880-4a14-a587-77331b1239b2","order_by":4,"name":"Jian Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+klEQVRIiWNgGAWjYDACCSB+wHCAgYG9h4EhAUkQv5aEBKAWnjMka5HIQRXECQxuNx/8kPjjjpz5zLcHPzzcYZdncID54G0eBrs8nFruHEuWSEh4ZixzOy9ZIvFMcrHBAbZkax6G5GJcWsxu5BgAtRxOnCENZCS2MSduOMBjJs3DcCCxAaeW/M8/gFrqZ0ieMf6R2FYP1ML/jYCWHDaQLQkSEjxmQFsOg2xhw6vF/kaamUVC2mHDGTw5ZhaJbccTZx5mM7acY5CMU4vkjOTHNz7YHJaXYD9jfPNnW3Vi3/HmhzfeVNjh1IIFMIMIA+LVj4JRMApGwSjABACraFyzeqs7WQAAAABJRU5ErkJggg==","orcid":"","institution":"Hainan University","correspondingAuthor":true,"prefix":"","firstName":"Jian","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2024-03-12 08:47:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4082073/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4082073/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":53744148,"identity":"590cf0db-9c5b-4a92-9ad5-d451895cd8b5","added_by":"auto","created_at":"2024-03-29 17:12:38","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":700015,"visible":true,"origin":"","legend":"\u003cp\u003eVisual comparisons on a real underwater image.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/a95e3cc7750528a52658dcf5.png"},{"id":53744150,"identity":"a356edb3-316f-41a6-b7da-044bf64f9787","added_by":"auto","created_at":"2024-03-29 17:12:38","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":530182,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the architecture of MFUNet.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/ebce31d08263e7ef1d88bfb3.png"},{"id":53744604,"identity":"ea26ca21-9f3b-491b-84ff-c3d92eccd863","added_by":"auto","created_at":"2024-03-29 17:20:38","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":478775,"visible":true,"origin":"","legend":"\u003cp\u003eRMT-guided multi-scale fusion process\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/77b9eef47d97d3649c16f166.png"},{"id":53744151,"identity":"82e680ff-78e8-485e-969c-b59e591456d3","added_by":"auto","created_at":"2024-03-29 17:12:38","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":49022,"visible":true,"origin":"","legend":"\u003cp\u003eSpecific details of the multi-scale fusion process\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/fff4751a35ac1d226ba641b5.png"},{"id":53744147,"identity":"d8cb1e2f-03b1-4291-a9e1-869bfbe8407f","added_by":"auto","created_at":"2024-03-29 17:12:38","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":34008,"visible":true,"origin":"","legend":"\u003cp\u003eWeighting of the medium transport map\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/f6847757d4419fd181018576.png"},{"id":53744152,"identity":"9b652106-2f0c-491f-acfb-74b62d03a0b8","added_by":"auto","created_at":"2024-03-29 17:12:39","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":48363,"visible":true,"origin":"","legend":"\u003cp\u003eChannel Attention\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/5a39c7508268a4e2e49a211f.png"},{"id":53744606,"identity":"25054ab0-9f65-4cdb-a199-d9609e263c8b","added_by":"auto","created_at":"2024-03-29 17:20:39","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":30136,"visible":true,"origin":"","legend":"\u003cp\u003eResidual Enhancement Module\u003c/p\u003e","description":"","filename":"image7.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/4323f08a48880572120587ce.png"},{"id":53744157,"identity":"c0e28ae1-5b54-4813-907c-443d036b4ca9","added_by":"auto","created_at":"2024-03-29 17:12:39","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":1119418,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of visual effects of images sampled by Test-S1000, the upper right corner of the image is the value of PSNR/MSE\u003c/p\u003e","description":"","filename":"image8.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/432d9e14a3719c7be74101aa.png"},{"id":53744155,"identity":"4f135dfa-9619-40d6-a511-2abeefbb7a7f","added_by":"auto","created_at":"2024-03-29 17:12:39","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":1600224,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of visual effects of images with obvious color cast and low contrast in Test-R90, the upper left corner of the image is the value of PSNR/MSE\u003c/p\u003e","description":"","filename":"image9.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/c362566b67f35bc498b0fbe4.png"},{"id":53744158,"identity":"79815581-7c7f-4001-bc29-70dba2fd18f5","added_by":"auto","created_at":"2024-03-29 17:12:39","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":721027,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of visual effects of images in Test-C60, the upper left corner of the image is the value of UIQM\u003c/p\u003e","description":"","filename":"image10.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/e1c2118e65fcff9c14f386fe.png"},{"id":53744605,"identity":"447cd362-61e6-4a88-8d07-020d27da6977","added_by":"auto","created_at":"2024-03-29 17:20:39","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":729767,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of visual effects of images in SQUID, the upper left corner is the value of UIQM\u003c/p\u003e","description":"","filename":"image11.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/def6ba36c4f69ff6fe42a85d.png"},{"id":53744153,"identity":"1cd66698-b802-4c9e-a88c-cfab210ac706","added_by":"auto","created_at":"2024-03-29 17:12:39","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":958403,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of visual effects of images in Color Checkerr, the upper left corner is the value of CIEDE2000\u003c/p\u003e","description":"","filename":"image12.png","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/94fb47ce5888a92a9501bb0a.png"},{"id":55594847,"identity":"7443d181-0162-4e46-aa7d-a8c25acb5115","added_by":"auto","created_at":"2024-04-30 10:20:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":10222161,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4082073/v1/44a17785-d1d1-409d-93ce-43f33d9ab81b.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Underwater Image Enhancement via Multi-Scale Feature Fusion Network Guided by Medium Transmission","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eEnhancing underwater images is a critical task for advancing marine applications and services, including underwater surveillance, image and video compression and transmission, and target detection. However, underwater imagery often suffers from significant color distortion, and the visuals are marred by low clarity and contrast due to the wavelength-dependent absorption of light, as well as scattering and refraction. Consequently, Improving the visual quality of underwater images through image enhancement is a formidable challenge.\u003c/p\u003e \u003cp\u003eTraditional underwater image enhancement methods mainly include non-physical model-based methods\u003csup\u003e1\u0026ndash;4\u003c/sup\u003e and physical model-based methods\u003csup\u003e5\u0026ndash;9\u003c/sup\u003e. Physical model-based methods\u003csup\u003e9\u0026ndash;15\u003c/sup\u003e focus on the accurate estimation of medium transmission and use the estimated medium transmittance and key underwater imaging parameters, such as uniform background light, to recover underwater images to obtain high quality images. They\u003csup\u003e16\u003c/sup\u003e proposed a Dark Channel Prior method for image dehazing, called DCP. The method obtains clear images by estimating atmospheric light and transmission maps. Due to its simplicity and effectiveness, many recovery-based UIE methods have modified the DCP to address the severe attenuation of red light in water. Chiang and Chen\u003csup\u003e10\u003c/sup\u003e used a defogging algorithm that compensates for light attenuation to enhance underwater images. Galdran et al.\u003csup\u003e17\u003c/sup\u003e proposed a Red-Channel Prior method (RCP) to recover the relevant colors at short wavelengths in order to restore contrast. Based on DCP, Peng et al.\u003csup\u003e18\u003c/sup\u003e proposed an Underwater DCP (UDCP) that only considers green and blue channels for underwater image restoration. The non-physical model method achieves the purpose of restoration by achieving contrast enhancement. They\u003csup\u003e2\u003c/sup\u003e proposed a method based on a non-physical model, called Histogram Equalization (HE), which uses pixel-value transformations to transform the original image into roughly the same number of pixels at most grey levels to recover the image. The limitations of traditional image methods mainly lie in ignoring the underwater imaging mechanism, resulting in insufficient or excessive enhancement, such as HE\u003csup\u003e2\u003c/sup\u003e, or being time-consuming or sensitive to the diversity of underwater scenes, such as DCP\u003csup\u003e16\u003c/sup\u003e, RCP\u003csup\u003e17\u003c/sup\u003e etc.\u003c/p\u003e \u003cp\u003eWith the development of deep learning technology, image enhancement methods based on deep learning are gradually becoming mainstream\u003csup\u003e19\u0026ndash;21\u003c/sup\u003e. The deep learning approach employs hierarchical feature extraction and original image reconstruction to produce high-quality restored images.The multi-color space encoder proposed by Ucolor\u003csup\u003e22\u003c/sup\u003e can integrate the features of different color spaces into a unified structure, improving the color space expression of images. However, it does not consider the feature fusion between different levels, does not pay attention to the structural similarity and scale difference of the restored image, and does not smooth the edges during the enhancement process, resulting in a large semantic gap between each feature layer and the loss of important details in the enhancement result. The U-Net\u003csup\u003e23\u003c/sup\u003e architecture introduces a fully symmetrical (U-shaped) network structure. It concatenates down-sampled and up-sampled feature maps of identical dimensions through skip connections, enabling the effective integration of high-level and low-level features. This method proves particularly beneficial in scenarios with limited sample sizes, offering rapid processing speeds. However, the U-Net\u003csup\u003e23\u003c/sup\u003e architecture falls short in representing features across different color spaces, and its feature depth is inadequate for leveraging the full potential of multi-scale feature fusion and multi-color space encoders.\u003c/p\u003e \u003cp\u003eTo address the issues identified, we leverage the strengths of both physical model-based and deep learning-based approaches. By integrating multi-scale fusion with enriched encoder features within the context of target detection tasks, we aim to correct color bias and amplify the contrast of underwater images. In our research, we employ a medium transfer map to quantify the extent of quality degradation in underwater images, facilitating targeted enhancement of deteriorated areas. And we use adaptive spatial fusion algorithms\u003csup\u003e24\u003c/sup\u003e to achieve feature interactions at non-adjacent levels through a progressive feature pyramid network and filter features in the multilevel fusion process so as to avoid large semantic gaps between non-adjacent layers, which allows us to retain useful information for fusion. Different from their method\u003csup\u003e22\u003c/sup\u003e, we use the medium transfer map to guide the feature pyramid, so that degraded regions can be enhanced when different feature layers interact. This method not only incorporates the underwater imaging mechanism into the network, but also enables the network to better retain underwater image details and improve the generalization ability of the network, thereby accelerating network optimization and improving performance. Inspired by their method\u003csup\u003e25\u003c/sup\u003e, we have incorporated an edge penalty term into the training regimen of our network. This adjustment heightens the network's sensitivity to edges and textures in underwater images. Our network is purely data-driven, designed to accommodate inaccuracies in media transfer estimates without compromising performance.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, we give a representative example, which is a comparison diagram between our proposed MFUNet and the other two underwater image enhancement methods. As shown in the figure, the fusion-based approach\u003csup\u003e26\u003c/sup\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e(b)) fails to repair the degraded regions and ignores them instead; the deep learning-based approach\u003csup\u003e22\u003c/sup\u003e(Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e(c)) handles the details poorly. Our MFUNet(Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e(d)) performs better in repairing degraded areas and enhancing details, and the repair results are more comfortable and natural. The main contributions of this article are as follows:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eMedium Transmission-Guided Multi-Scale Fusion Network: We introduce a medium transmission-guided multi-scale fusion network, where features in the multilevel fusion process undergo adaptive spatial fusion operations. This strategy prevents the loss or degradation of information throughout the multilevel transmission process, ensuring the retention of valuable information. Additionally, by utilizing a medium transmission map, we direct the network's attention towards regions of quality degradation, thereby harmonizing the strengths of both physical model-based and deep learning-based approaches.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eEdge Loss Function Implementation: Our methodology incorporates an edge loss function by introducing an edge penalty term during the network's training phase. This compels the network to heighten its sensitivity towards the edges and textures of underwater images. Consequently, this enhances the network's responsiveness to high-frequency information, leading to an improvement in image quality.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eEnhanced Network Encoding and Decoding Levels: By deepening the encoding and decoding levels of our network, we capitalize on the benefits of the adaptive spatial fusion algorithm and the multi-color encoder. This amplifies the network's feature representation capabilities and boosts overall performance.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ePerformance Benchmarking: Our MFUNet achieves unparalleled performance in terms of objective metrics when compared to contemporary benchmark methods.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this section, we discuss related work on underwater image enhancement. Underwater image enhancement methods can be roughly divided into: methods based on physical models, methods without physical models, and methods based on deep learning.\u003c/p\u003e \u003cp\u003eThe methods without physical model directly modifies the image pixel values to improve visibility, such as multi-scale fusion\u003csup\u003e26\u0026ndash;28\u003c/sup\u003e, variational optimization\u003csup\u003e29\u003c/sup\u003e, and pixel distribution adjustment \u003csup\u003e30\u003c/sup\u003e, etc. For example, Ancuti et al. \u003csup\u003e26\u003c/sup\u003e first obtain the color correction and contrast enhancement maps of an underwater image respectively; then calculate the corresponding weight maps; and finally combine the respective advantages to obtain the enhancement maps. Further, Ancuti et al. found that the information contained in at least one-color channel is completely lost in unfavorable scenes such as nighttime, underwater and uneven artificial illumination. So they proposed a color channel compensation preprocessing method to improve the traditional restoration method. The physical model-free methods ignore the underwater imaging mechanism, which tends to produce excessive or insufficient results and introduces artificial colors that produce unstable enhancement results.\u003c/p\u003e \u003cp\u003eThe physical model-based approach\u003csup\u003e12,17,18,31,32\u003c/sup\u003etreats single underwater image enhancement as a pathological inverse problem, where the parameters of the underwater image formation model are estimated by manual a priori. The prior includes underwater dark channel prior\u003csup\u003e31\u003c/sup\u003e, red channel prior \u003csup\u003e17\u003c/sup\u003e, minimum informati\u003csup\u003e12\u003c/sup\u003eon prior and fuzzy prior\u003csup\u003e12\u003c/sup\u003e, etc. These methods take into account the scattering and attenuation of light and achieve good results. Akkaynak and Treibitz \u003csup\u003e33\u003c/sup\u003e proposed a model that uses modified underwater images to form a model with physical accuracy. And based on the revised model, a new underwater image color correction method\u003csup\u003e34\u003c/sup\u003e based on RGB-D image pairs was proposed. Peng and Cosman\u003csup\u003e32\u003c/sup\u003e proposed an underwater image depth estimation algorithm based on image blur and light absorption, which uses the estimated depth and underwater imaging model to enhance underwater images. Peng et al.\u003csup\u003e35\u003c/sup\u003e further proposed generalizing the dark channel before processing different images captured under severe weather. However, in challenging underwater conditions, statistical priors may fail, and underwater image formation models may produce unreliable enhancement results\u003csup\u003e33,36\u003c/sup\u003e due to different scene properties. Moreover, this method may require manual correction and is time-consuming.\u003c/p\u003e \u003cp\u003eDeep learning-based methods have made significant progress in underwater image enhancement\u003csup\u003e20\u0026ndash;22,37-42\u003c/sup\u003e. Due to the lack of underwater images and corresponding clean image pairs, previous works have used generative adversarial networks (Generative Adversarial Network, GAN) to synthesize underwater image pairs or perform solution pair learning. Li et al. \u003csup\u003e20\u003c/sup\u003e first used a Generative Adversarial Network (GAN) to synthesize degraded images and proposed a two-stage refinement network. To meet the requirement for paired training data. \u003csup\u003e19\u003c/sup\u003e proposes a weakly supervised underwater color correction network (UCcycleGAN). Uplavikar et al. \u003csup\u003e38\u003c/sup\u003e introduced a simple classifier to make the GAN model more discriminative for different water types. In order to solve the problem of lack of unpaired data sets in supervised learning \u003csup\u003e21\u003c/sup\u003e, real underwater images were simulated based on the underwater imaging physical model and 10 different water types\u003csup\u003e43\u003c/sup\u003e. And they proposed a lightweight CNN model trained on each of the ten different water types and an underwater image enhancement benchmark based on real-world underwater image pairs respectively (UIEB) \u003csup\u003e41\u003c/sup\u003e. PRWnet further proposes a gate fusion network that fuses three enhanced inputs to obtain enhanced underwater images. \u003csup\u003e22\u003c/sup\u003e proposes a multi-color space embedding network that combines the advantages of physical models to deal with color casts and low contrast in underwater images. Existing underwater image enhancement methods mostly focus on the spatial domain, and it is difficult to recover high and low frequency mixed distortion images,\u003csup\u003e25\u003c/sup\u003e proposes a wavelet enhancement-based learning network to gradually refine underwater images in the spatial and frequency domains. The current deep learning-based methods do not focus on the structural similarity and image quality of the recovered images, ignore the targets with scale differences, and the network hierarchy is shallow, which cannot give full play to the advantages of each module.\u003c/p\u003e \u003cp\u003eCompared with existing deep learning-based underwater image enhancement methods, our approach has the following advantages:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eWe introduce an adaptive spatial fusion algorithm and guide it through a medium transmission map, and we integrate the underwater imaging principle into the deep structure so that the restored image can adaptively retain the original image information, continuously enhancing the degraded area.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe calculate the edge loss using the Laplace operator and make it serve as a penalty term during the network training process, enabling the network to restore the high-frequency information of the image.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWe have deepened the depth of the network, so that each module of the network can play a better performance.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eOur method achieves excellent results on a variety of underwater image datasets.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"3. Proposed Method","content":"\u003cp\u003eIn this section, the proposed MFUNet framework is introduced in detail. It includes Overall Architecture(3.1),RMT-guided multi-scale fusion module(3.2), network layer deepening(3.3), multi-color space encoder(3.4), channel attention module(3.5), residual enhancement module(3.6) and loss function(3.7).\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1.Network structure\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe have designed a multi-color space encoder network based on multi-scale fusion(MFUnet), and Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the general architecture of the network.\u003c/p\u003e \u003cp\u003eIn this network, the input underwater image is first subjected to color space transformation, so that the input image enters three color coding paths: RGB path, Lab path and HSV path. In each color path, the input image is delivered to four serial enhancement modules. Each serial enhancement module performs a 2-fold down-sampling operation on the input to obtain four levels of feature representation. The three-color paths are tightly fused with the corresponding features of the RGB path, the HSV path and the Lab path while the serial enhancement is performed. The RGB paths are enhanced by replacing the original features of the RGB channels with the fused features. Subsequently, three parallel identical channel level features are concatenated, resulting in four sets of multi-color space encoder features, represented as [C2,C3,C4,C5]. And these four sets of color features are fed into the channel attention module of the corresponding inlet, which serves to extract the most representative semantic features. After obtaining the [C2,C3,C4,C5], they are input into the multi-scale fusion module guided by the medium transfer map (RMT). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the extraction and fusion between multilevel features is achieved. The specific way to implement this is that the bottom features C2 and C3 are first fed to the feature pyramid. Then the features are weighted using an RMT map of the same size to compensate for quality degradation regions, denoted as [C2`, C3`]. Then feature fusion is performed on [C2`, C3`] to obtain a set of outputs. Then C4 is added and finally C5 is added to get the final multi-scale features weighted by the RMT map, represented as [P2, P3, P4, P5]. In the feature pyramid, RMT maps of different scales are implemented by maximum pooling. Ordinary convolution kernel is used for feature up-sampling and bilinear interpolation is used for feature down-sampling. Finally, the [P2,P3,P4,P5] of the output of the multi-scale fusion module are fed to the corresponding residual enhancement modules. After four serial residual modules with three times 2-fold up-sampling, the resulting decoded features are fed to the convolutional layer for the resultant reconstruction, resulting in an enhanced image.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2.RMT-guided asymptotic feature pyramid network\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows our proposed RMT-guided asymptotic feature pyramid network (R-AFPN) process.\u003c/p\u003e \u003cp\u003eTo enhance the clarity and richness of details and textures in the resultant underwater images, while suppressing noise and artifacts, and improving regions of image degradation, we utilize the Reverse-Asymptotic Feature Pyramid Network (R-AFPN) module for feature fusion across different scales. Given that lower-layer features possess higher resolution but smaller receptive fields, and upper-layer features have larger receptive fields but lower resolution, we employ a hierarchical, multi-layer fusion strategy. This approach enriches the receptive fields of all-layer features, allowing each layer to assimilate semantic information from others, thereby bridging the semantic gap and boosting the overall network performance. However, a challenge arises during multi-scale fusion: the degraded regions within underwater images tend to merge, potentially exacerbating image degradation. To address this, we apply Reverse Medium Transmission (RMT) maps at varying scales to pre-compensate for feature degradation prior to fusion at each layer. This solution effectively slove the issue, enhancing the network's sensitivity to degraded areas and overall performance. Using RMT for weighted compensation is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eSince the semantic gap between non-adjacent hierarchical features is larger than the semantic gap between adjacent hierarchical features, especially the gap between the bottom and the top, it directly leads to poor fusion of non-adjacent hierarchical features. Therefore it is not reasonable to use [C2,C3,C4,C5] directly for feature fusion. Therefore, we adopt progressive multi-scale fusion, which makes the semantic information of different level features closer together again in the process of progressive fusion, thereby alleviating the above problems.\u003c/p\u003e \u003cp\u003eOur approach is to extract the last layer of each feature layer of multi-color channel encoding to obtain a set of feature [C2,C3,C4,C5] of different scales. In the bottom-up feature extraction process of the backbone network, the R-AFPN module gradually performs RMT degradation compensation on the bottom, high-level and top-level features and fuses them with each other. Specifically, the bottom features are first compensated for RMT degradation and fused with the bottom features to obtain new bottom features. Then the depth and the new bottom features are compensated for RMT degradation and fused, and the same is done for the top features. We use 1\u0026times;1 convolution and bilinear interpolation methods to up-sample features, and use different convolution kernels and steps to perform down-sampling. After each feature fusion is completed, a residual block is used for feature learning for each feature layer. Each residual block consists of two 3\u0026times;3 convolutions. For the specific fusion process, we assign different spatial weights to the features at different levels in the multilevel feature fusion process. In this way, we enhance the importance of key levels and reduce the influence of interference information to achieve the purpose of self-adaptation. The specific details are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe image formation model\u003csup\u003e44,45\u003c/sup\u003e under severe weather is widely used in image dehazing and underwater image restoration algorithms\u003csup\u003e10,32,35\u003c/sup\u003e, and its quality degradation image can be expressed as:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ1\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$${\\text{I}}^{\\text{c}}\\left(\\text{x}\\right)={\\text{J}}^{\\text{c}}\\left(\\text{x}\\right)\\otimes \\text{T}\\left(\\text{x}\\right)\\oplus {\\text{A}}^{\\text{c}}\\left(\\text{x}\\right)\\otimes \\left(1-\\text{T}\\left(\\text{x}\\right)\\right),\\text{c}\\in \\{\\text{r},\\text{g},\\text{b}\\},$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn the formula, x represents the pixel index, I is the observation image, J is the clear image, A is the uniform background light, and T is the medium projection rate, which means the percentage of scene brightness that reaches the camera after being reflected in the medium, reflecting the quality degradation in different areas degree.\u003c/p\u003e \u003cp\u003eWe incorporate the medium transmission map into the multi-scale fusion network via the proposed medium transmission guidance module. Specifically, the reverse medium transport (RMT)map,represented as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\overline{\\text{T}}\\in {\\mathbb{ℝ}}^{\\text{H}\\times \\text{W}}\\)\u003c/span\u003e\u003c/span\u003e,is used as a pixel-level attention map. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\overline{\\text{T}}\\)\u003c/span\u003e\u003c/span\u003e in the RMT diagram is obtained from 1-\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{T}\\)\u003c/span\u003e\u003c/span\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\((\\text{T}\\in {\\mathbb{ℝ}}^{\\text{H}\\times \\text{W}}.\\text{I}\\text{t} \\text{i}\\text{s} \\text{r}\\text{e}\\text{p}\\text{r}\\text{e}\\text{s}\\text{e}\\text{n}\\text{t}\\text{s} \\text{t}\\text{h}\\text{e} \\text{T} \\text{i}\\text{s} [\\text{0,1}]\\)\u003c/span\u003e\u003c/span\u003erange media transfer map\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(, \\text{a}\\text{n}\\text{d} 1\\in {\\mathbb{ℝ}}^{\\text{H}\\times \\text{W}}\\)\u003c/span\u003e\u003c/span\u003e represents matrices with all elements 1) This means that more degraded areas should be allocated greater attention.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eHowever, real medium transport maps corresponding to underwater images are lacking in practice, so the depth model cannot be trained for medium transport map estimation. Inspired by robust general dark channel prior\u003csup\u003e35\u003c/sup\u003e, we estimate the medium transport map as:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ2\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\stackrel{\\sim}{\\text{T}}\\left(\\text{x}\\right)={\\text{m}\\text{a}\\text{x}}_{\\text{c},\\text{y}\\in {\\Omega }\\left(\\text{x}\\right)}\\left(\\frac{{\\text{A}}^{\\text{c}}-{\\text{I}}^{\\text{c}}\\left(\\text{y}\\right)}{\\text{m}\\text{a}\\text{x}\\left({\\text{A}}^{\\text{c}},1-{\\text{A}}^{\\text{c}}\\right)}\\right),$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\stackrel{\\sim}{\\text{T}}\\)\u003c/span\u003e\u003c/span\u003e is the estimated medium transmission map, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\Omega }\\left(\\text{x}\\right)\\)\u003c/span\u003e\u003c/span\u003e represents a local patch with a size of 15 \u0026times; 15 centered on x, and c denotes the color channe, and A represents the uniform background light. In the method\u003csup\u003e35\u003c/sup\u003e, the estimation of uniform background light is based on depth-dependent color changes.\u003c/p\u003e \u003cp\u003eBased on the RMT map, the principle of the proposed medium transport bootstrap module is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. As can be seen from Eq. We use the RMT map as a feature selector to weight the different spatial locations of the features. Pixels with high degradation (pixels with larger RMT values) will be given higher weights, which can be expressed as:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ3\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\mathcal{V}=\\mathcal{U}\\oplus \\mathcal{U}\\otimes \\overline{\\text{T}},$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\mathcal{V}\\in {\\mathbb{ℝ}}^{\\text{M}\\times \\text{H}\\times \\text{W}} \\text{a}\\text{n}\\text{d} \\mathcal{U}\\in {\\mathbb{ℝ}}^{\\text{M}\\times \\text{H}\\times \\text{W}}\\)\u003c/span\u003e\u003c/span\u003e in the formula represent the output features and input features of the media transmission guidance module respectively. We treat the RMT weight as an identity connection to avoid vanishing gradients and tolerate errors caused by inaccurate estimation of medium transmission. Furthermore, our purely data-driven framework is also tolerant of inaccuracies in the media transfer graph.\u003c/p\u003e \u003cp\u003eInspired by the method\u003csup\u003e24\u003c/sup\u003e, we use an adaptive spatial fusion algorithm to enhance the importance of key levels and mitigate the effect of contradictory information from different objects. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, we fused three levels of features. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{x}}_{\\text{i}\\text{j}}^{\\text{n}\\to \\text{l}}\\)\u003c/span\u003e\u003c/span\u003erepresents the feature vectors at the level from \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{n}\\)\u003c/span\u003e\u003c/span\u003e level to \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{l} \\text{l}\\text{e}\\text{v}\\text{e}\\text{l}\\left(\\text{i},\\text{j}\\right)\\)\u003c/span\u003e\u003c/span\u003e. The \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{y}}_{\\text{i}\\text{j}}^{\\text{l}},\\)\u003c/span\u003e\u003c/span\u003e feature vector obtained by adaptive spatial fusion of multilevel features is obtained by \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{x}}_{\\text{i}\\text{j}}^{1\\to \\text{l}},{\\text{x}}_{\\text{i}\\text{j}}^{2\\to \\text{l}},{\\text{x}}_{\\text{i}\\text{j}}^{3\\to \\text{l}}\\)\u003c/span\u003e\u003c/span\u003e linear combination. It is shown below:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ4\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$${\\text{y}}_{\\text{i}\\text{j}}^{\\text{l}}={{\\alpha }}_{\\text{i}\\text{j}}^{\\text{l}}\\cdot {\\text{x}}_{\\text{i}\\text{j}}^{1\\to \\text{l}}+{{\\beta }}_{\\text{i}\\text{j}}^{\\text{l}}\\cdot {\\text{x}}_{\\text{i}\\text{j}}^{2\\to \\text{l}}+{{\\gamma }}_{\\text{i}\\text{j}}^{\\text{l}}\\cdot {\\text{x}}_{\\text{i}\\text{j}}^{3\\to \\text{l}},$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn the formula, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({{\\alpha }}_{\\text{i}\\text{j}}^{\\text{l}},{{\\beta }}_{\\text{i}\\text{j}}^{\\text{l}}\\text{与}{{\\gamma }}_{\\text{i}\\text{j}}^{\\text{l}}\\)\u003c/span\u003e\u003c/span\u003e represents the spatial weight of the three-level features of level \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{l}\\)\u003c/span\u003e\u003c/span\u003e, subject to the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({{\\alpha }}_{\\text{i}\\text{j}}^{\\text{l}}+{{\\beta }}_{\\text{i}\\text{j}}^{\\text{l}}+{{\\gamma }}_{\\text{i}\\text{j}}^{\\text{l}}=1\\)\u003c/span\u003e\u003c/span\u003e constraint.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3.Deepening of the network hierarchy\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe structure of U-net\u003csup\u003e23\u003c/sup\u003e network facilitates the integration of information from the various stages of down-sampling in the up-sampling process, so that the high level of the network obtains the high-frequency information of the graph, and the bottom level of the network obtains the low-frequency information of the graph. That is, the up-sampling process combines the structural information of the various levels. The information at each level is then preserved using jump connections. The traditional network structure\u003csup\u003e23\u003c/sup\u003e is usually a three-layer network structure, but after our experiments, we learned that deepening the network layers can further play the advantages of multi-color encoder to enhance the color spatial representation of the image, and play the advantages of multi-scale fusion to reduce the semantic gap of each level through increasing the sensitivity of the restored image to the degraded region, and improve the model's ability to generalize and enhance the performance of the image. The results are shown in the ablation experiment.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4.Multi-color space encoder\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAccording to the underwater imaging model, images captured underwater exhibit significant color deviations, often appearing bluer, greener, or yellower than those observed on land.\u003c/p\u003e \u003cp\u003eThe diverse color spectrum of underwater images often renders traditional network architectures ineffective. To address this challenge, it is essential to adopt a novel network architecture that ensures enhanced underwater images appear more natural. Inspired by some multi-color enhancement algorithms\u003csup\u003e22\u003c/sup\u003e working in color spaces, an image will have different expressions in different color spaces, so we extracted the features from three color spaces of the image, RGB, HSV and Lab, in order to correct the color deviation of the underwater image.\u003c/p\u003e \u003cp\u003eCurrently, the RGB color space is the most prevalent, primarily because the human eye predominantly perceives the three colors of red, green, and blue, making operations within the RGB color space relatively straightforward and amenable to mathematical processing. However, the components of R, G, and B are highly interrelated and significantly influenced by variations in brightness and shadow, complicating color correction through other methods.In the HSV color space, the hue, saturation, brightness and contrast of the image can be intuitively reflected. In the Lab color space, all colors that can be observed by the human eye can be displayed. The characteristics of each color space are very obvious, and the correlation between color spaces is also low, which provides prerequisites for color correction of underwater images.\u003c/p\u003e \u003cp\u003eWe integrate features from various color spaces into a cohesive depth structure, enabling the utilization of components such as color, hue, and saturation, which are pertinent to image degradation. Importantly, the color disparity between two adjacent points might be negligible within one color space, yet pronounced in others. Therefore, the embedding of multi-color space makes it easier to measure the color deviation of an underwater image and correct the image. In addition, the multi-color encoder brings more non-linear operations in the color space transformation process, which significantly improves the performance of the depth model.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.5.Channel attention module\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eGiven the distinct definitions of each color space, it is crucial to assign appropriate weights to the features extracted from the three color spaces to optimize the functionality of the multi-color encoder. To achieve this, we utilize a channel attention module to illustrate the interdependence among different color spaces. The channel attention module is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSuppose the input feature is \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\begin{array}{r}\\mathcal{ℱ}=\\mathsf{C}\\mathsf{a}\\mathsf{t}\\left({\\text{F}}_{1},{\\text{F}}_{2},\\cdots ,{\\text{F}}_{\\text{N}}\\right)\\in \\end{array}{ \\mathbb{ℝ}}^{\\text{N}\\times \\text{H}\\times \\text{W}}\\)\u003c/span\u003e\u003c/span\u003e,F is the feature mapping of a path at a particular level; N is the number of feature mappings; Cat is the feature stitching; H and W are the height and width of the input image, respectively. We first perform global average pooling on the input features F to obtain the channel descriptor \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{z}\\in {\\mathbb{ℝ}}^{\\text{N}\\times 1}\\)\u003c/span\u003e\u003c/span\u003e,followed by the embedded global distribution of the channel feature responses.The k-th item of z can be expressed as:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ5\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\begin{array}{r}{\\text{z}}_{\\text{k}}=\\frac{1}{\\text{H}\\times \\text{W}}\\sum _{\\text{i}}^{\\text{H}}\\sum _{\\text{j}}^{\\text{W}}{\\text{F}}_{\\text{k}}\\left(\\text{i},\\text{j}\\right),\\end{array}$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn the formula \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{k}\\in \\left[1,\\text{N}\\right]\\)\u003c/span\u003e\u003c/span\u003e. In order to fully extract the dependencies of each channel, we used self-gating mechanism\u003csup\u003e46\u003c/sup\u003e to obtain the set of modulation weights for each channel \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\text{s}\\in {\\mathbb{ℝ}}^{\\text{N}\\times 1}\\)\u003c/span\u003e\u003c/span\u003e:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ6\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e\n$$\\text{s}={\\sigma }\\left({\\text{W}}_{2}\\ast \\left({\\delta }\\left({\\text{W}}_{1}\\ast \\text{z}\\right)\\right)\\right),$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\sigma }\\left(\\cdot \\right)\\)\u003c/span\u003e\u003c/span\u003e represents the Sigmoid activation function, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\delta }\\left(\\cdot \\right)\\)\u003c/span\u003e\u003c/span\u003e represents the ReLU activation function, * represents the convolution operation, W1 and W2 are the weights of the two fully connected layers respectively, and their output channel numbers are \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\frac{\\text{N}}{\\text{r}}\\)\u003c/span\u003e\u003c/span\u003e and N,respectively, where r is 16, in order to reduce the computational cost. Finally, these weights will be applied to the input features F to generate the rescaled features \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\mathcal{U}\\in {\\mathbb{ℝ}}^{\\text{N}}\\times \\text{H}\\times \\text{W}\\)\u003c/span\u003e\u003c/span\u003e.Furthermore, in order to maintain the good properties of the original features while avoiding the problem of gradient vanishing, we treat the channel-attention weights in the same mapping manner:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ7\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ7\" name=\"EquationSource\"\u003e\n$$\\mathcal{U}=\\mathcal{ℱ}\\oplus \\mathcal{ℱ}\\otimes \\text{s},$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e7\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003ewhere\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\oplus\\)\u003c/span\u003e\u003c/span\u003eand\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\otimes\\)\u003c/span\u003e\u003c/span\u003edenote pixel-by-pixel addition and pixel-by-pixel multiplication, respectively.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.6.Residual enhancement module\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e shows the residual enhancement module we used. The primary objective of the residual enhancement module is to ensure the integrity and fidelity of the data, while also preventing the occurrence of gradient explosion or vanishing. To enhance the overall performance and stability of the network, we strategically place residual enhancement modules throughout the network, at each up-sampling and down-sampling layer. Within these modules, the convolutional layers are standardized to have an identical number of filters. In each residual enhancement module, the convolutional layers have the same number of filters. In the encoder network, the number of filters is gradually increased from 16 to 128 by a factor of 2. In the decoder network, the number of filters is reduced from 128 to 16 by a factor of 2. All the convolutional layers have a convolutional kernel of 3 \u0026times; 3 and a step size of 1.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.7.Loss function\u003c/h2\u003e \u003cp\u003eWe use linear combination of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{l}}_{2}\\)\u003c/span\u003e\u003c/span\u003e loss \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{{l}_{2}}\\)\u003c/span\u003e\u003c/span\u003e, perceptual loss \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{L}}_{\\text{p}\\text{e}\\text{r}}\\)\u003c/span\u003e\u003c/span\u003e and edge loss \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{L}}_{\\text{e}\\text{d}\\text{g}}\\)\u003c/span\u003e\u003c/span\u003e to balance visual quality and quantitative scores:\u003cdiv id=\"Equ8\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ8\" name=\"EquationSource\"\u003e\n$$L={\\lambda }_{1}{L}_{{l}_{2}}+{\\lambda }_{2}{L}_{per}+{\\lambda }_{3}{L}_{edg}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e8\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{{l}_{2}}\\)\u003c/span\u003e\u003c/span\u003erepresents the loss \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({l}_{2}\\)\u003c/span\u003e\u003c/span\u003e; \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{per}\\)\u003c/span\u003e\u003c/span\u003e represents the perceptual loss; \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{edg}\\)\u003c/span\u003e\u003c/span\u003e represents the edge loss;\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\lambda }_{1}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\lambda }_{2}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\lambda }_{3}\\)\u003c/span\u003e\u003c/span\u003e are the balancing factors. In this chapter let \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\lambda }_{1}\\)\u003c/span\u003e\u003c/span\u003e=5, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\lambda }_{3}\\)\u003c/span\u003e\u003c/span\u003e=0.05, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\lambda }_{3}\\)\u003c/span\u003e\u003c/span\u003e=10. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({l}_{2}\\)\u003c/span\u003e\u003c/span\u003e loss \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{{l}_{2}}\\)\u003c/span\u003e\u003c/span\u003e measures the difference between the reference image \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(I\\)\u003c/span\u003e\u003c/span\u003e and the reconstructed image :\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ9\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ9\" name=\"EquationSource\"\u003e\n$${L}_{{l}_{2}}=\\sum _{m=1}^{H}\\sum _{n=1}^{W}{\\left(\\widehat{I}\\left(m,n\\right)-I\\left(m,n\\right)\\right)}^{2}$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e9\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn the formula, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(I\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\widehat{I}\\)\u003c/span\u003e\u003c/span\u003e are the reference and reconstructed images, respectively; \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(H\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(W\\)\u003c/span\u003e\u003c/span\u003e are the height and width of the images. The perceptual loss \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{per}\\)\u003c/span\u003e\u003c/span\u003e combines the features extracted from the VGG-19\u003csup\u003e47\u003c/sup\u003e network to measure the structural consistency of the reconstructed image and the reference image. Perceptual loss is defined as follows:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ10\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ10\" name=\"EquationSource\"\u003e\n$${L}_{per}=\\sum _{m=1}^{H}\\sum _{n=1}^{W}\\left|{\\varphi }_{i}\\left(\\widehat{I}\\right)\\left(m,n\\right)-{\\varphi }_{i}\\left(I\\right)\\left(m,n\\right)\\right|$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e10\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn the formula, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\varphi\\)\u003c/span\u003e\u003c/span\u003e represents the VGG-19 network pre-trained on the ImagNet dataset, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\varphi }_{i}\\left(\\cdot \\right)\\)\u003c/span\u003e\u003c/span\u003e is the i-th convolutional layer. Here we use the features output by the relu5_4 layer of the VGG-19 network. The edge loss \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({L}_{edg}\\)\u003c/span\u003e\u003c/span\u003e is used to reconstruct the edges and texture of the image and the edge loss is denoted as:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Equ11\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ11\" name=\"EquationSource\"\u003e\n$${L}_{edg}=\\sqrt{\\parallel \\varDelta \\widehat{I}-\\varDelta I{\\parallel }^{2}+{\\epsilon }^{2}}$$\u003c/div\u003e \u003cdiv class=\"EquationNumber\"\u003e11\u003c/div\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\varDelta\\)\u003c/span\u003e\u003c/span\u003e denotes the Laplace operator and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\epsilon\\)\u003c/span\u003e\u003c/span\u003e is set to \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({10}^{-3}\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Experiments","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eNext, the implementation details of this chapter are described first, and then the software and hardware environment and experimental settings of the experiment are introduced. We compared the method proposed in this chapter with representative methods and performed ablation experiments to verify the effectiveness of each part.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.1.Data sets\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo be fair, our training set and benchmarks are consistent with the literature\u003csup\u003e41\u003c/sup\u003e. They are introduced separately below.\u003c/p\u003e \u003cp\u003e(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e)Training set. To train the algorithms proposed in this chapter, we randomly selected 800 pairs of underwater images from the UIEB\u003csup\u003e41\u003c/sup\u003e underwater image enhancement dataset. The UIEB dataset consists of 890 pairs of degraded underwater pictures and the corresponding reference versions, which cover different underwater scenes and different degradation situations. We also selected 1250 images from a synthetic underwater dataset\u003csup\u003e21\u003c/sup\u003e, divided into subsets of 10 water quality types, including open seawater as I, IA, IB, II, III, and coastal waters 1, 3, 5, 7, and 9. Additionally, to increase the training data, we randomly cropped the patches with a size of 128x128.\u003c/p\u003e \u003cp\u003e(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e)Benchmarks. We use the remaining 90 pairs of images from the UIEB dataset, denoted as Test-R90. we select 100 pairs of images from each subset of the synthetic underwater dataset\u003csup\u003e21\u003c/sup\u003e, denoted as Test-R1000. In addition, we also conduct comprehensive experiments on Test-C60\u003csup\u003e41\u003c/sup\u003e, SQUID\u003csup\u003e48\u003c/sup\u003e and Color-Check7\u003csup\u003e27\u003c/sup\u003e. Test-C60 contains 60 real underwater images that do not have the reference images provided in the UIEB. Test-C60 is more challenging compared to Test-R90. The Squud dataset contains 57 pairs of underwater images from 4 different dive sites in Israel. We used 16 representative examples provided on the SQUID project page for testing. Specifically, 4 representative samples from each of the 4 dive sites (Katzaa, Michmoret, Nachsholim, Satil) were selected, and the resolution of each image was 1827\u0026times;2737.Color-Check7 contains 7 underwater color checker images taken with different cameras provided in the literature, and was used to evaluate the underwater robustness and accuracy of color correction. The cameras used to take the color-check photographs are denoted in this paper as Can D10, Fuj Z33, Oly T6000, Oly T8000, Pen W60, Pen W80, and Pan TS1.\u003c/p\u003e \u003cp\u003e(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e)Comparison of methods. We compare with 11 other typical underwater image enhancement methods. These include a non-physical modelling approach (Ancutiet al\u003csup\u003e26\u003c/sup\u003e.), three physical model-based approaches, (Li et al\u003csup\u003e12\u003c/sup\u003e., Peng et al\u003csup\u003e32\u003c/sup\u003e., GDCP\u003csup\u003e35\u003c/sup\u003e), four learned approaches, (UcycleGAN\u003csup\u003e19\u003c/sup\u003e, Guo et al\u003csup\u003e49\u003c/sup\u003e., Water-Net\u003csup\u003e20\u003c/sup\u003e,UWCNN\u003csup\u003e21\u003c/sup\u003e) and three baseline depth models(Unet-U\u003csup\u003e23\u003c/sup\u003e、Unet-RMT\u003csup\u003e23\u003c/sup\u003e and Ucolor\u003csup\u003e22\u003c/sup\u003e). Due to hardware platform limitations, both the methods proposed in this chapter and Ucolor training reduce the batch size to 10. This undoubtedly reduces the actual measured performance data, but does not affect the conclusions of the final experiments. The data for the other compared methods are taken from the publicly available data in literature\u003csup\u003e41\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.2.Evaluation metrics.\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFor Test-R90 and Test-S1000, we used PSNR, SSIM and MSE for evaluation. Similar to the literature\u003csup\u003e41\u003c/sup\u003e, the reference image of Test-R90 is used as the real image in order to calculate the PSNR, SSIM and MSE scores. A higher PSNR score or a higher SSIM score or a lower MSE score indicates that the resultant image is closer to the image content of the real image and the algorithm performs better.\u003c/p\u003e \u003cp\u003eFor Test-C60 and SQUID that do not have corresponding ground truth images, we employ the no-reference evaluation metrics UCIQE\u003csup\u003e50\u003c/sup\u003e and UIQM\u003csup\u003e51\u003c/sup\u003e to measure the performance of different methods. The Higher UCIQE or UIQM scores indicate that the image is visually better. In addition, we also provide NIQE\u003csup\u003e52\u003c/sup\u003e test scores. The lower NIQE scores indicate better image quality. Due to hardware limitations, we scaled the images in the SQUID test set to 512\u0026times;512 and then processed them with different algorithms.\u003c/p\u003e \u003cp\u003eFor Color-Check7, we tested the difference in swatch colors between the real image and the corresponding enhanced image. Specifically, we cropped 24 color blocks from the enhancement result and the corresponding real image and calculated the average color value of each block separately. Then the method in literature\u003csup\u003e27\u003c/sup\u003e is adopted, and CIEDE\u003csup\u003e53\u003c/sup\u003e is used to measure the relative perceptual difference between corresponding color patches. The smaller the value of CIEDE, the better the image enhancement effect.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.3.Implementation details\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe software environment we used is Python 3.6, TensorFlow 1.9, cudnn 7.1.4, cuda10.0. The hardware platform is Intel Xeon E5-2680 for CPU, 32GB of RAM, and Titan Xp for GPU with 12GB of video memory. The batch size for training is 10, and the training period is 500. The channel base is 16, initialized with Gaussian distribution, the optimizer is Adam, the learning rate is fixed to 1e-4 and the training time is one day.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.4.Visual comparisons\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this section, we conduct a visual comparison across different test sets, beginning with an evaluation of the visualization effects on synthetic images.\u003c/p\u003e \u003cp\u003eAs depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, the algorithm introduced in this chapter outperforms the competing algorithms in terms of image dehazing and color correction. Among all the algorithms assessed, the one proposed in this study most closely approximates the scene structure of the actual image, achieving the highest Peak Signal-to-Noise Ratio (PSNR) ,and the lowest Mean Squared Error (MSE) values.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e, we show the results of a real underwater image with a significant green color bias sampled from Test-R90 processed by different algorithms. The color bias in the degraded underwater image obscures the details of the image, and correcting the color bias results in clearer details. In terms of color correction, all the comparison methods are unsatisfactory, with the colors showing either under or over compensation. In particular, GDCP and Ancuti introduce additional color artefacts. In contrast, the method proposed in this chapter effectively eliminates color bias, improves contrast, and without obvious under-compensation, over-compensation or artifacts.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe also sampled challenging images from Test-C60 for comparison, which generally suffer from severe degradation such as color shift, blurring, low brightness, etc., and the results of different methods of enhancement are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e. In all the results showcased, none of the compared methods achieved satisfactory outcomes; some algorithms produced artifacts, while others introduced artificial colors. The highest Underwater Image Quality Measure (UIQM) was achieved by the algorithm proposed in this study.\u003c/p\u003e \u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e, we show the visual effect of sampled images in SQUID. Ancuti et al.'s algorithm achieved the best contrast, but the image showed a green color cast. The method proposed in this chapter can effectively improve the contrast of the image without producing obvious artificial color shift.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe have compared the processing results of underwater Color Checker images and the results are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e. All the compared methods try to correct the color deviation as a whole, but the proposed method in this paper has the best color correction results.\u003c/p\u003e \u003cp\u003eThe comparisons of visual effects presented above demonstrate that the algorithm introduced in this study outperforms all other algorithms in terms of visual quality. Furthermore, it exhibits robustness across various datasets and shows excellent generalizability to different underwater scenes.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.5.Quantitative comparisons\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn order to further verify the performance of the algorithm proposed in this chapter, we first conducted a quantitative comparison on the Test-S1000 and Test-R90 data sets. Due to hardware platform limitations, we change the training batch size of Ucolor to 10. The average PSNR, MSE (\u0026times;1000) and SSIM scores of different methods are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, and the best results for each indicator are marked in red. As shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the performance of the algorithm proposed in this chapter is better than almost all compared methods. Compared with Ucolor, which ranks second in overall performance, the PSNR, MSE and SSIM scores on Test-S1000 are increased by 7.3%, 44.5% and 44.5% respectively. 3.0%. The PSNR, MSE and SSIM scores on Test-R90 are improved by 1.3%, 5.5% and 3.9% respectively.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation results of different algorithms on TEST-S1000 and TEST-R90, \"-\" indicates that the results are not available, and \u0026ldquo;*\u0026rdquo; indicates that the training batch is 10.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMethods\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eTest-S1000\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eTest-R90\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePSNR\u0026uarr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMSE\u0026darr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSSIM\u0026uarr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePSNR\u0026uarr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMSE\u0026darr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSSIM\u0026uarr;\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003einput\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAncuti et al\u003csup\u003e26\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e19.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLi et al\u003csup\u003e13\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14.29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeng et al\u003csup\u003e32\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e15.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGDCP\u003csup\u003e35\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e11.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3.40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.50\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGuo et al\u003csup\u003e49\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e15.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e18.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUcycleGAN\u003csup\u003e19\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWater-Net \u003csup\u003e41\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e15.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e19.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN type1\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN type3\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e15.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN type5\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN type7\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN type9\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN typeI\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e10.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e10.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e6.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN typeII\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e14.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN typeIII\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e3.40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN retrain21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e15.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnet-U\u003csup\u003e23\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e19.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e18.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnet-RMT\u003csup\u003e23\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e17.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUcolor*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e20.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e20.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOurs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e21.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e20.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe tested the performance of different algorithms on the Test-C60 and SQUID datasets. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the average scores of UIQM, UCIOE and NIQE for different algorithms. The best results for each are marked in red. It can be seen that none of the algorithms outperforms the others on these two challenging datasets. Overall, Li et al.'s algorithm performs the best, and the performance of the algorithms proposed in this chapter is in the middle of the pack overall.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eUIQM, UCIQE and NIQE scores of different algorithms on Test-C60 and SQUID datasets. \"-\" indicates that the result is not available and \u0026ldquo;*\u0026rdquo; indicates a training batch of 10.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMethods\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eTest-C60\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eSQUID\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUIQM\u0026uarr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUCIQE\u0026uarr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNIQE\u0026darr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eUIQM\u0026uarr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eUCIQE\u0026uarr;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNIQE\u0026darr;\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003einput\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e7.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAncuti et al\u003csup\u003e26\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e4.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5.01\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLi et al\u003csup\u003e13\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.81\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeng et al\u003csup\u003e32\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.39\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGDCP\u003csup\u003e35\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGuo et al\u003csup\u003e49\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUcycleGAN\u003csup\u003e19\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e7.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5.93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWater-Net\u003csup\u003e41\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.72\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN typeII\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.60\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN retrain\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.38\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnet-U\u003csup\u003e23\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.38\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnet-RMT\u003csup\u003e23\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5.16\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUcolor*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e7.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e7.50\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOurs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.096\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e7.26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eNext, we compared the underwater check colour card images in Color-Check7 to demonstrate the accuracy and robustness of the colour correction of the method proposed in this chapter, and the results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The best results for each are marked in red. The performance of the algorithm proposed in this chapter is optimal in most cases (6 cameras), and only slightly worse than its competitors on Pen W80 and Pan TS1.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCIEDE2000(\u0026darr;) scores of different algorithms on Color-Check7, \u0026ldquo;*\u0026rdquo; indicates the training batch is 10\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMethods\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePen W60\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePen W80\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCan D10\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFuj Z33\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eOly T6000\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eOly T8000\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003ePan TS1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eAvg\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003einput\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e16.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e14.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e23.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e19.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e17.24\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAncuti et al\u003csup\u003e26\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e13.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e14.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e11.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e12.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e10.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e12.32\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLi et al\u003csup\u003e13\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e15.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e18.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e25.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e16.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e17.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e12.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e17.38\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeng et al\u003csup\u003e32\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e14.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e14.09\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e12.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e14.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e19.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e14.95\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGDCP\u003csup\u003e35\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e15.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e24.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e16.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e12.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e16.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e12.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e16.13\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGuo et al\u003csup\u003e49\u003c/sup\u003e.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12.29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e15.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e14.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e39.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e15.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e12.40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e18.04\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUcycleGAN\u003csup\u003e19\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e21.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e21.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e22.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e26.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e20.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e23.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e19.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e22.14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWater-Net\u003csup\u003e41\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e19.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e15.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e17.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e21.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e18.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e16.94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN typeII\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e20.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e17.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e17.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e16.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e17.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e20.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e18.20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUWCNN retrain\u003csup\u003e21\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e20.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e14.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e13.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e14.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e18.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e20.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e16.47\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnet-U\u003csup\u003e23\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e11.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e15.17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e13.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e10.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e15.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e17.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e13.56\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnet-RMT\u003csup\u003e23\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e19.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e15.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e14.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e13.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e16.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e19.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e15.86\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUcolor*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e10.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e12.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e11.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e13.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e12.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e11.26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOurs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e10.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e7.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e11.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e13.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e10.27\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.6.Ablation study\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn order to analyze the effectiveness of each core component of the method proposed in this chapter, we conducted an ablation study on the Test-R90 data set. It includes RMT-guided multi-scale feature fusion (R-AFPN), 4-level refined network structure (4-layers) and edge loss (edge-loss). As shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the complete model proposed in this chapter achieved the best results, demonstrating the effectiveness of the method proposed in this article. The results for each of the other core component combination methods also confirmed the necessity of each component.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eresults of ablation experiments on Test-R90\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eR-AFPN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4-layers\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eedge-loss\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePSNR\u003c/p\u003e \u003cp\u003e(\u0026uarr;)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMSE\u0026times;10\u0026sup3;\u003c/p\u003e \u003cp\u003e(\u0026darr;)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSSIM\u003c/p\u003e \u003cp\u003e(\u0026uarr;)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUIQM\u003c/p\u003e \u003cp\u003e(\u0026uarr;)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eUCIQE\u003c/p\u003e \u003cp\u003e(\u0026uarr;)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eNIQE\u003c/p\u003e \u003cp\u003e(\u0026darr;)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e20.005\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.968\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.745\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.385\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.558\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e7.996\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e20.356\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.832\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.742\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.422\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.564\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e7.742\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e19.974\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.955\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.766\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.461\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.566\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e7.617\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e19.965\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.976\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.745\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.441\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.561\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e7.197\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e20.413\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.862\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.773\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.441\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.567\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e7.834\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e20.037\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.929\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.767\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.404\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.563\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e7.331\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026times;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e19.971\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.917\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.745\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.337\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.554\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e7.416\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026radic;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e20.439\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.862\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.767\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e1.542\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.571\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e7.410\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.7.Comparison with Ucolor\u003csup\u003e22\u003c/sup\u003e\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo further exemplify our improvements, Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows a comparison of the model size and other performances of our proposed network with the Ucolor network. In order to make the network work on embedded platforms with limited hardware, we reduced the training parameters of the network, achieved by changing the channel base from 128 to 16, which reduces the enhancement effect but in return results in a lightweight network. In the comparison of model size, Ours is reduced by 73.05% compared to the Ucolor network. In terms of training time, Ours takes only 1/3 of the time required by the Ucolor network, and is 8.89% faster than the Ucolor network in terms of average speed of image processing.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison with the Ucolor network, \u0026ldquo;*\u0026rdquo; indicates the training batch is 10*, \u0026ldquo;C60\u0026rdquo;indicates testing on the Test-C60 dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCompare items\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUcolor*\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOurs\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e756.09MB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e203.78MB\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraining duration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThree days\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOne days\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eProcessing speed\u003csup\u003eC60\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e171.0s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e155.8s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChannel base\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe introduce an underwater image enhancement method leveraging multi-scale feature fusion. Traditional approaches fail to address the issue of information loss and degradation due to the indirect interaction between non-adjacent layers, nor do they fully harness the potential of each module. Moreover, most existing image enhancement techniques focus on spatial domain enhancement, neglecting frequency domain information enhancement. Our method employs a progressive feature pyramid network to overcome these challenges. Specifically, we adopt progressive strategies and self-adaptation for feature fusion. During the fusion process, the medium transmission map is utilized as a weight of importance, integrating underwater imaging domain knowledge into the network. We enhance the network's coding and decoding depth to maximize the benefits of each module and employ edge loss as a penalty term to amplify frequency domain detail enhancement. Extensive testing on various benchmarks confirms the superiority of our approach.\u003c/p\u003e \u003cp\u003eHowever, our method exhibits certain limitations. Firstly, the enhancement of frequency domain details is primarily focused on edge loss, resulting in suboptimal performance in enhancing frequency domain details. Secondly, the estimation of the medium transmission map, which guides the multi-scale fusion, is occasionally inaccurate, thereby impacting the overall enhancement effect. To address these issues, future efforts will concentrate on augmenting the frequency domain enhancement techniques and refining the guidance method to further improve our approach.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe UIEB dataset is available at https://li-chongyi.github.io/proj_benchmark.html with credentialed access.\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003eThis work was supported by the College Student Innovation and Entrepreneurship Project of Hainan\u0026nbsp;University(Hdcxcyxm201704), the Hainan Provincial Natural Science Foundation of China (623RC449).\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003eY.H: Conceptualization, supervision, methodology, software, validation, formal analysis, writing-original draft, writing-review\u0026amp;edition;\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003eC.H.Q.: conceptualization, methodology, data curation, software, validation, formal analysis, writing-original draft, writing-review\u0026amp;edition; J.C.X.: data curation, resources, writing-review\u0026amp;edition, visualization;\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003eZ.R.T.: resources, funding acquisition;\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003eZ.J.: supervision, writing-review\u0026amp;edition, project administration, funding acquisition;\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003eAll authors read and approved the final article.\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp dir=\"LTR\"\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eGao, S. B., Zhang, M., Zhao, Q., Zhang, X. S. \u0026amp; Li, Y. J. Underwater Image Enhancement Using Adaptive Retinal Mechanisms. IEEE Trans Image Process 28, 5580\u0026ndash;5595, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2019.2919947\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2019.2919947\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSinghai, J. \u0026amp; Rawat, P. in \u003cem\u003eInternational Conference on Computational Intelligence and Multimedia Applications (ICCIMA 2007).\u003c/em\u003e 507\u0026ndash;512 (IEEE).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, W., Dong, L., Zhang, T. \u0026amp; Xu, W. Enhancing underwater image via color correction and bi-interval contrast enhancement. Signal Processing: Image Communication 90, 116030, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/vcip.2017.8305027\u003c/span\u003e\u003cspan address=\"10.1109/vcip.2017.8305027\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhuang, P. \u0026amp; Ding, X. Underwater image enhancement using an edge-preserving filtering retinex algorithm. Multimedia Tools Applications 79, 17257\u0026ndash;17277, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s11042-020-08739-3\u003c/span\u003e\u003cspan address=\"10.1007/s11042-020-08739-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDrews, P., Nascimento, E., Moraes, F., Botelho, S. \u0026amp; Campos, M. in \u003cem\u003eProceedings of the IEEE international conference on computer vision workshops.\u003c/em\u003e 825\u0026ndash;830.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSong, W., Wang, Y., Huang, D. \u0026amp; Tjondronegoro, D. in \u003cem\u003eAdvances in Multimedia Information Processing\u0026ndash;PCM 2018: 19th Pacific-Rim Conference on Multimedia, Hefei, China, September 21\u0026ndash;22\u003c/em\u003e, 2018, \u003cem\u003eProceedings, Part I 19.\u003c/em\u003e 678\u0026ndash;688 (Springer).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSong, W., Wang, Y., Huang, D., Liotta, A. \u0026amp; Perra, C. Enhancement of underwater images with statistical model of background light and optimization of transmission map. IEEE Transactions on Broadcasting 66, 153\u0026ndash;169, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TBC.2019.2960942\u003c/span\u003e\u003cspan address=\"10.1109/TBC.2019.2960942\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu, H. \u003cem\u003eet al.\u003c/em\u003e Underwater image enhancement based on DCP and depth transmission map. Multimedia Tools and Applications 79, 20373\u0026ndash;20390, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s11042-020-08701-3\u003c/span\u003e\u003cspan address=\"10.1007/s11042-020-08701-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhuang, P., Li, C. \u0026amp; Wu, J. Bayesian retinex underwater image enhancement. Engineering Applications of Artificial Intelligence 101, 104171, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.engappai.2021.104171\u003c/span\u003e\u003cspan address=\"10.1016/j.engappai.2021.104171\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChiang, J. Y. \u0026amp; Chen, Y.-C. Underwater image enhancement by wavelength compensation and dehazing. \u003cem\u003eIEEE transactions on image processing\u003c/em\u003e 21, 1756\u0026ndash;1769, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2011.2179666\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2011.2179666\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDrews, P. L., Nascimento, E. R., Botelho, S. S. \u0026amp; Campos, M. F. M. Underwater depth estimation and image restoration based on single images. IEEE computer graphics \u0026amp; applications 36, 24\u0026ndash;35, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/MCG.2016.26\u003c/span\u003e\u003cspan address=\"10.1109/MCG.2016.26\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, C.-Y., Guo, J.-C., Cong, R.-M., Pang, Y.-W. \u0026amp; Wang, B. Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior. IEEE Transactions on Image Processing 25, 5664\u0026ndash;5677, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2016.2612882\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2016.2612882\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, C. \u003cem\u003eet al.\u003c/em\u003e in \u003cem\u003eIEEE International Conference on Image Processing (ICIP).\u003c/em\u003e 1993\u0026ndash;1997 (IEEE).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerman, D., Treibitz, T. \u0026amp; Avidan, S. in \u003cem\u003eProc. British Machine Vision Conference (BMVC).\u003c/em\u003e 2.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, C., Guo, J., Guo, C., Cong, R. \u0026amp; Gong, J. A hybrid method for underwater image correction. Pattern Recognition Letters 94, 62\u0026ndash;67, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.patrec.2017.05.023\u003c/span\u003e\u003cspan address=\"10.1016/j.patrec.2017.05.023\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, K., Sun, J. \u0026amp; Tang, X. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis 33, 2341\u0026ndash;2353, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TPAMI.2010.168\u003c/span\u003e\u003cspan address=\"10.1109/TPAMI.2010.168\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaldran, A., Pardo, D., Pic\u0026oacute;n, A. \u0026amp; Alvarez-Gila, A. Automatic red-channel underwater image restoration. Journal of Visual Communication and Image Representation 26, 132\u0026ndash;145, doi:10.1016/j.jvcir.2014.11.006. (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDrews, P. L., Nascimento, E. R., Botelho, S. S., Campos, M. F. M. J. I. c. g. \u0026amp; applications. Underwater depth estimation and image restoration based on single images. 36, 24\u0026ndash;35, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/MCG.2016.26\u003c/span\u003e\u003cspan address=\"10.1109/MCG.2016.26\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, C., Guo, J. \u0026amp; Guo, C. Emerging from water: Underwater image color correction based on weakly supervised color transfer. IEEE Signal processing letters 25, 323\u0026ndash;327, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/LSP.2018.2792050\u003c/span\u003e\u003cspan address=\"10.1109/LSP.2018.2792050\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, J., Skinner, K. A., Eustice, R. M. \u0026amp; Johnson-Roberson, M. WaterGAN: Unsupervised generative network to enable real-time color correction of monocular underwater images. IEEE Robotics and Automation letters 3, 387\u0026ndash;394, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/LRA.2017.2730363\u003c/span\u003e\u003cspan address=\"10.1109/LRA.2017.2730363\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, C., Anwar, S. \u0026amp; Porikli, F. Underwater scene prior inspired deep underwater image and video enhancement. Pattern Recognition 98, 107038, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.patcog.2019.107038\u003c/span\u003e\u003cspan address=\"10.1016/j.patcog.2019.107038\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, C. \u003cem\u003eet al.\u003c/em\u003e Underwater image enhancement via medium transmission-guided multi-color space embedding. IEEE Transactions on Image Processing 30, 4985\u0026ndash;5000, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2021.3076367\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2021.3076367\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRonneberger, O., Fischer, P. \u0026amp; Brox, T. in \u003cem\u003eMedical Image Computing and Computer-Assisted Intervention\u0026ndash;MICCAI 2015: 18th International Conference, Munich, Germany, October 5\u0026ndash;9\u003c/em\u003e, 2015, \u003cem\u003eProceedings, Part III 18.\u003c/em\u003e 234\u0026ndash;241 (Springer).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, G. \u003cem\u003eet al.\u003c/em\u003e in \u003cem\u003eIEEE International Conference on Systems, Man, and Cybernetics (SMC).\u003c/em\u003e 2184\u0026ndash;2189 (IEEE).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuo, F., Li, B. \u0026amp; Zhu, X. in \u003cem\u003eProceedings of the IEEE/CVF International Conference on Computer Vision.\u003c/em\u003e 1944\u0026ndash;1952.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAncuti, C., Ancuti, C. O., Haber, T. \u0026amp; Bekaert, P. in \u003cem\u003eIEEE conference on computer vision and pattern recognition.\u003c/em\u003e 81\u0026ndash;88 (IEEE).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAncuti, C. O., Ancuti, C., De Vleeschouwer, C. \u0026amp; Bekaert, P. Color balance and fusion for underwater image enhancement. IEEE Transactions on image processing 27, 379\u0026ndash;393, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2017.2759252\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2017.2759252\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAncuti, C. O., Ancuti, C., Haber, T. \u0026amp; Bekaert, P. in \u003cem\u003eIEEE International Conference on Image Processing.\u003c/em\u003e 1557\u0026ndash;1560 (IEEE).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFu, X. \u003cem\u003eet al.\u003c/em\u003e in \u003cem\u003e2014 IEEE international conference on image processing (ICIP).\u003c/em\u003e 4572\u0026ndash;4576 (IEEE).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhani, A. S. A. \u0026amp; Isa, N. A. M. Underwater image quality enhancement through integrated color model with Rayleigh distribution. Applied soft computing 27, 219\u0026ndash;230, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.asoc.2014.11.020\u003c/span\u003e\u003cspan address=\"10.1016/j.asoc.2014.11.020\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDrews, P. L., Nascimento, E. R., Botelho, S. S. \u0026amp; Campos, M. F. M. Underwater depth estimation and image restoration based on single images. IEEE computer graphics applications 36, 24\u0026ndash;35, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/MCG.2016.26\u003c/span\u003e\u003cspan address=\"10.1109/MCG.2016.26\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeng, Y.-T. \u0026amp; Cosman, P. C. Underwater image restoration based on image blurriness and light absorption. IEEE transactions on image processing 26, 1579\u0026ndash;1594, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2017.2663846\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2017.2663846\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAkkaynak, D. \u0026amp; Treibitz, T. in \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition.\u003c/em\u003e 6723\u0026ndash;6732.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAkkaynak, D. \u0026amp; Treibitz, T. in \u003cem\u003eProceedings of the IEEE/CVF conference on computer vision and pattern recognition.\u003c/em\u003e 1682\u0026ndash;1691.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeng, Y.-T., Cao, K. \u0026amp; Cosman, P. C. Generalization of the dark channel prior for single image restoration. IEEE Transactions on Image Processing 27, 2856\u0026ndash;2868, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2018.2813092\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2018.2813092\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAkkaynak, D. \u003cem\u003eet al.\u003c/em\u003e in \u003cem\u003eProceedings of the IEEE Conference on Computer Vision and Pattern Recognition.\u003c/em\u003e 4931\u0026ndash;4940.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFabbri, C., Islam, M. J. \u0026amp; Sattar, J. in 2018 \u003cem\u003eIEEE international conference on robotics and automation (ICRA).\u003c/em\u003e 7159\u0026ndash;7165 (IEEE).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUplavikar, P. M., Wu, Z. \u0026amp; Wang, Z. in \u003cem\u003eCVPR workshops.\u003c/em\u003e 1\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJamadandi, A. \u0026amp; Mudenagudi, U. in \u003cem\u003eProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops.\u003c/em\u003e 11\u0026ndash;17.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIslam, M. J., Xia, Y., Sattar, J. J. I. R. \u0026amp; Letters, A. Fast underwater image enhancement for improved visual perception. 5, 3227\u0026ndash;3234, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/LRA.2020.2974710\u003c/span\u003e\u003cspan address=\"10.1109/LRA.2020.2974710\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, C. \u003cem\u003eet al.\u003c/em\u003e An underwater image enhancement benchmark dataset and beyond. IEEE Transactions on Image Processing 29, 4376\u0026ndash;4389, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2019.2955241\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2019.2955241\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJamadandi, A. \u0026amp; Mudenagudi, U. in \u003cem\u003eCVPR Workshops.\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJerlov, N. G. \u003cem\u003eMarine optics\u003c/em\u003e. 1, 2 (Elsevier, 1976).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNarasimhan, S. G. \u0026amp; Nayar, S. K. Vision and the atmosphere. International journal of computer vision 48, 233\u0026ndash;254, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/1508044.1508113\u003c/span\u003e\u003cspan address=\"10.1145/1508044.1508113\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (2002).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan, R. T. in \u003cem\u003eIEEE conference on computer vision and pattern recognition.\u003c/em\u003e 1\u0026ndash;8 (IEEE).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu, J., Shen, L. \u0026amp; Sun, G. in \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition.\u003c/em\u003e 7132\u0026ndash;7141.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimonyan, K. \u0026amp; Zisserman, A. Very deep convolutional networks for large-scale image recognition. Computer Science, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.1409.1556\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1409.1556\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerman, D., Levy, D., Avidan, S. \u0026amp; Treibitz, T. Underwater single image color restoration using haze-lines and a new quantitative dataset. IEEE transactions on pattern analysis and machine intelligence 43, 2822\u0026ndash;2837, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TPAMI.2020.2977624\u003c/span\u003e\u003cspan address=\"10.1109/TPAMI.2020.2977624\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo, Y., Li, H. \u0026amp; Zhuang, P. Underwater image enhancement using a multiscale dense generative adversarial network. IEEE Journal of Oceanic Engineering 45, 862\u0026ndash;870, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/JOE.2019.2911447\u003c/span\u003e\u003cspan address=\"10.1109/JOE.2019.2911447\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, M. \u0026amp; Sowmya, A. An underwater color image quality evaluation metric. IEEE Transactions on Image Processing 24, 6062\u0026ndash;6071, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TIP.2015.2491020\u003c/span\u003e\u003cspan address=\"10.1109/TIP.2015.2491020\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePanetta, K., Gao, C. \u0026amp; Agaian, S. Human-visual-system-inspired underwater image quality measures. IEEE Journal of Oceanic Engineering 41, 541\u0026ndash;551, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/JOE.2015.2469915\u003c/span\u003e\u003cspan address=\"10.1109/JOE.2015.2469915\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMittal, A., Soundararajan, R. \u0026amp; Bovik, A. C. Making a \u0026ldquo;completely blind\u0026rdquo; image quality analyzer. IEEE Signal processing letters 20, 209\u0026ndash;212, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/LSP.2012.2227726\u003c/span\u003e\u003cspan address=\"10.1109/LSP.2012.2227726\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSharma, G., Wu, W. \u0026amp; Dalal, E. N. The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research \u0026amp; Application 30, 21\u0026ndash;30, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/col.20070\u003c/span\u003e\u003cspan address=\"10.1002/col.20070\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2005).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Image enhancement, Underwater imaging, Deep learning, Multi-scale feature fusion","lastPublishedDoi":"10.21203/rs.3.rs-4082073/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4082073/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDue to the complexity of underwater imaging environments, images captured via optical vision systems often exhibit significant degradation. To combat this issue, we introduce a multi-scale feature fusion underwater image enhancement network, termed MFUNet. MFUNet is a novel multi-scale feature fusion network, guided by medium transmission, ensures the content integrity of the reconstructed image by leveraging interaction features among non-adjacent layers. This approach addresses the common problem of the loss of image detail features. Moreover, MFUNet enhances the response to high-frequency information by employing edge loss, thereby improving sensitivity to edges and textures. By deepening the network hierarchy, the image undergoes deep encoding and decoding, which maximizes the multi-color space encoder's and multi-scale feature fusion's potential in color representation and enhances the structural similarity and overall quality of the image. It is worth noting that we achieved superior performance by utilizing fewer model parameters. Extensive experiments across various datasets demonstrate that our method surpasses comparative methods in both visual quality and quantitative metrics.\u003c/p\u003e","manuscriptTitle":"Underwater Image Enhancement via Multi-Scale Feature Fusion Network Guided by Medium Transmission","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-03-29 17:12:34","doi":"10.21203/rs.3.rs-4082073/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"82ef1c46-a76e-4bd6-adc5-a2bda0d6e06c","owner":[],"postedDate":"March 29th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":29895717,"name":"Physical sciences/Mathematics and computing/Computational science"},{"id":29895718,"name":"Physical sciences/Mathematics and computing/Computer science"},{"id":29895719,"name":"Physical sciences/Mathematics and computing/Scientific data"}],"tags":[],"updatedAt":"2024-04-30T10:11:36+00:00","versionOfRecord":[],"versionCreatedAt":"2024-03-29 17:12:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4082073","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4082073","identity":"rs-4082073","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}