Bridging the Gap Between Low-Cost Cameras and High-Fidelity Monitoring: Deployment of Super-Resolution Models for Real-World Lettuce Farming

Bridging the Gap Between Low-Cost Cameras and High-Fidelity Monitoring: Deployment of Super-Resolution Models for Real-World Lettuce Farming | 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 Research Article Bridging the Gap Between Low-Cost Cameras and High-Fidelity Monitoring: Deployment of Super-Resolution Models for Real-World Lettuce Farming Mhd. Idham Khalif, Tjhwa Endang Djuana, Richard Antonius Rambung, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9305065/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 Currently, the application of visual monitoring in smart agriculture is one of the options in the implementation of precision agriculture. Smart agriculture visual monitoring has challenges in terms of the relatively high cost of high-resolution cameras and limited access to resources. One option that can be used is implementing embedded low-resolution cameras so that the cost is also low and improves the quality of image resolution by implementing a deep learning-based Super-Resolution (SR) method. This study applies image enhancement to low-resolution cameras directly using three deep learning-based SR models—EDSR, Real-ESRGAN, and ESPCN—for a 2× resolution increase from 800 × 600 (SVGA) to 1600 × 1200 (UXGA) to find out the SR model that is suitable for precision agriculture according to the conditions. Experiments were conducted using a real-world lettuce growth dataset taken by ESP32-CAM as input to the low-resolution camera and implemented on NVIDIA Jetson Orin Nano as edge computing. Performance was assessed in terms of reconstruction quality, computational load, processing latency, and power consumption under CPU and GPU execution. The results show that Real-ESRGAN achieves the highest visual quality at the expense of computational and energy requirements, EDSR offers a good balance, and ESPCN provides the highest efficiency with reduced image detail. These findings highlight the potential for low-cost visual growth monitoring of lettuce plants under limited resource constraints, leading to precision agriculture applications. super-resolution edge computing smart agriculture deep learning visual monitoring 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 Plant growth monitoring is a fundamental component of modern agricultural systems, enabling continuous observation of plant development, health assessment, and early detection of productivity-threatening disturbances [ 1 ]. Advances in digital technology have accelerated the adoption of smart agriculture, where Artificial Intelligence (AI) plays a key role in improving efficiency and crop yields. Camera-based image monitoring enables continuous and non-invasive data acquisition without manual intervention [ 2 , 3 ]. However, high-quality imaging systems often involve high costs and energy consumption, limiting their scalability in resource-constrained environments [ 4 – 6 ]. The ESP32-CAM has emerged as a low-cost alternative due to its compact size, wireless connectivity, and suitability for edge computing applications [ 7 ]. Nevertheless, its limited image resolution can negatively affect downstream tasks such as plant detection and analysis. Deep learning-based super-resolution (SR) techniques offer a promising solution by reconstructing high-resolution images from low-resolution inputs without increasing hardware cost. Although SR models have demonstrated strong performance [ 8 , 9 ], most are trained on synthetic datasets that do not reflect real-world degradation [ 10 ]. This study applies to edge computing, three SR architectures—EDSR [ 11 ], Real-ESRGAN [ 12 ], and ESPCN [ 13 ], using real lettuce image dataset captured by ESP32-CAM, so that it can be evaluated on the feasibility of implementation in edge-based vision system, so that it can be applied to low-cost visual precision lettuce crop monitoring system with optimal resource usage. 2. Related Work 2.1. Vision-Based Plant Growth Monitoring Various studies have applied computer vision-based technology to monitor plant growth in real-time. One early study utilized a Raspberry Pi-based camera to non-invasively monitor corn cob growth over a long period of time, and showed that the image-based system was able to capture plant growth dynamics on an ongoing basis [ 14 ]. Other research has developed a plant growth monitoring system using a web camera connected to a computer, where the image results are processed to model plant structure in three dimensions (3D), thus enabling more in-depth analysis of plant morphology [ 15 ]. Apart from that, the use of drones equipped with cameras has also been widely applied to monitor plant growth from an aerial perspective, especially in large-scale agricultural areas [ 16 ]. Although this camera-based approach has proven effective in monitoring plant growth, most of the systems developed still rely on camera devices with high specifications, both in terms of image resolution and processing capabilities. This dependency generally results in increased initial investment costs and system energy consumption. These conditions pose significant challenges to the implementation of crop monitoring systems that are sustainable, cost-effective and widely applicable, especially in agricultural environments with limited resources. 2.2. Super-Resolution (SR) for Image Enhancement SR techniques have been widely applied as a method for improving image quality (image enhancement) in various application domains. In the field of biomedical imaging, the Real-ESRGAN method is used to increase the resolution of tissue microscopic images, where low resolution images with a wide field of view (FOV) are successfully reconstructed into high resolution images without losing FOV, while expanding the depth of field (DOF) to a certain extent [ 17 ]. Another approach uses Efficient Sub-Pixel Convolutional Network (ESPCN) to increase the resolution of X-ray microscopic images, with evaluation results showing a significant increase in image quality and an average Peak Signal-to-Noise Ratio (PSNR) value reaching 40 dB [ 18 ]. In addition, Enhanced Deep Super Resolution (EDSR) has also been applied to three-dimensional (3D) X-ray images of porous media, resulting in image reconstruction with sharper details and clearer structures, and proving the effectiveness of EDSR in complex scientific imaging domains [ 19 ]. 2.3. Research Gap and Contribution Previous studies show that although computer vision and deep learning-based super-resolution (SR) techniques have been widely applied in various imaging fields, research evaluating SR model performance on plant images captured directly using low-cost cameras under real conditions remains limited. Most studies rely on synthetic datasets or high-specification imaging devices, which do not fully represent the real-world degradation characteristics of cameras such as ESP32-CAM. Moreover, comparative analyses that consider not only reconstructed image quality but also training efficiency and the potential implementation in cost- and energy-efficient crop monitoring systems—particularly for lettuce—are still scarce. Therefore, this study contributes by providing a comprehensive evaluation of three deep learning-based SR models—EDSR, Real-ESRGAN, and ESPCN—using lettuce image datasets captured directly from ESP32-CAM. The evaluation considers image quality, training dynamics, and deployment potential on edge computing, aiming to support the selection of appropriate SR models for low-cost camera-based plant monitoring systems in smart precision agriculture. 3. Methodology 3.1. Data Acquisition Setup In this study, the image data acquisition process was carried out using an ESP32-CAM device as the main image capture system, thus representing the real conditions of continuous lettuce plant growth monitoring over time. The selection of the ESP32-CAM was based on its characteristics as a low-cost camera widely used in Internet of Things (IoT) applications and smart agriculture systems. Despite having limitations in sensor resolution, dynamic range, and image quality stability, this device reflects the real challenges faced in implementing low-cost camera-based crop monitoring systems. Therefore, the resulting dataset has the characteristics of low-resolution images with realistic visual degradation, making it very relevant to evaluate the effectiveness of deep learning-based SR methods in improving crop image quality without the need for expensive hardware. Figure 1 illustrates the use of the ESP32-CAM as the image acquisition device in this study. Figure 1 (a) shows the placement of the ESP32-CAM on the growing medium to continuously document mustard plant growth during the observation period, while Fig. 1 (b) presents the front view of the camera setup. The camera was positioned at a fixed angle and distance from the plant to ensure consistent data acquisition. Images were captured every 30 minutes to record visual changes in plant appearance due to growth and environmental variations during both day and night. In each session, the ESP32-CAM sequentially captured two images with a one-second interval: an 800 × 600 pixel (SVGA) image as the low-resolution (LR) dataset and a 1600 × 1200 pixel (UXGA) image as the high-resolution (HR) dataset. SVGA resolution was selected as the LR representation because it provides a balance between visual detail and device limitations. Compared with VGA or lower resolutions, SVGA retains sufficient morphological information of plants, making it suitable for developing and evaluating SR techniques on low-cost cameras. Although the ESP32-CAM supports UXGA resolution, it is not always stable in practice, and some captured images may be corrupted compared to those at SVGA resolution, as illustrated in Fig. 2 . 3.2. Dataset Construction The dataset used in this study consists of a total of 2,505 low-resolution (LR) and high-resolution (HR) image pairs. All image pairs were acquired through the same data acquisition process and under identical capture conditions, so each LR image has a spatially and temporally corresponding HR image pair. This approach, similar to paired datasets captured in the real world, has been shown to be important in SR research, as DL models rely on accurate mapping between LR and HR images to learn features and detect degradation, and because model performance often degrades when trained solely on synthetic images that do not reflect real-world conditions [ 20 ]. This approach ensures that visual differences between LR and HR images are caused solely by differences in sensor resolution, not by external factors such as lighting changes, object movement, or variations in camera angle. Therefore, the resulting dataset is wellsuited for training and evaluating deep learning-based SR methods, as it allows for an objective and representative assessment of model performance against real-world conditions in low-cost crop monitoring systems. Figure 3 shows an example of a digital image dataset of lettuce plant growth used in this study. In general, the dataset was collected under two main lighting conditions: daytime, with adequate natural light intensity, and night time, with minimal ambient lighting and the aid of a flash. These varying lighting conditions are intended to represent real-world image acquisition conditions. In the dataset used, low-resolution (LR) and high-resolution (HR) image pairs served as the basis for evaluating the performance of a SR-based deep learning model. This evaluation aimed to assess the model’s ability to improve the quality of low-resolution images to approximate high-resolution reference images. The use of a relatively large paired dataset is expected to yield a more representative and reliable evaluation of model performance in real-world scenarios. Each LR–HR image pair was obtained from the same image capture session at two consecutive resolutions. This approach aims to ensure that the image pairs originate from identical acquisition conditions, ensuring that differences in image quality are solely due to differences in sensor resolution, rather than external factors such as changes in lighting or object movement. 3.3. Super-Resolution (SR) Methods 3.3.1. Enhanced Deep Super-Resolution (EDSR) Is a deep learning-based SR technique designed to improve the quality of low-resolution images into high-resolution images by utilizing Convolutional Neural Networks (CNNs) which are known for their good accuracy. EDSR eliminates non-essential components in conventional SR architectures, such as batch normalization, thus improving feature representation capabilities and producing image reconstructions with sharper details and higher accuracy [ 21 ]. Figure 4 , the Enhanced Deep Super-Resolution (EDSR) architecture [ 21 ], for 2x SR, consists of several key components designed to effectively reconstruct high-resolution (HR) images from low-resolution (LR) images. EDSR is trained using an end-to-end deep learning approach that allows feature learning to be performed directly from low-resolution images to high-resolution ones. This approach results in superior reconstruction performance, making EDSR known as one of the best-performing CNN-based architectures in SR tasks [ 22 , 23 ]. The process begins with an initial convolution layer that extracts basic features from the input image. These features are then processed through a series of residual blocks without batch normalization (ResBlock), as shown in Fig. 4 , to enhance feature representation while preserving the range of pixel intensity values. Each residual block employs a skip connection to maintain training stability and retain important information from previous layers. After passing through the residual blocks, the extracted features are combined with the initial features through a global residual connection before entering the upsampling stage. At this stage, EDSR applies a sub-pixel convolution (pixel shuffle) mechanism, illustrated in the Up sample section of Fig. 4 , to efficiently increase spatial resolution. This architectural design enables EDSR to produce sharper image reconstructions with 2× enhancement, making it suitable for improving low-resolution lettuce images captured using low-cost cameras such as ESP32-CAM. 3.3.2. Real-Enhanced Super-Resolution Generative Adversial Network (Real-ESRGAN) Is one of the methods in SR known as perceptual quality, to improve the visual quality of images that experience complex degradation in the real world through a generative model. This model is trained with synthetic data that simulates real degradation so that it is able to improve visual details and reduce artifacts better than previous methods on real datasets [ 24 ]. Figure 5 is the Real-ESRGAN architecture, the SR process begins with a pixel-unshuffle stage on low-resolution input images, especially for magnification scales of ×2 and ×1. This operation functions to reverse the pixel-shuffle mechanism by reducing the spatial resolution of the image while increasing the number of feature channels, so that the computational load can be reduced without losing important information. The pixel-unshuffled features are then processed through an initial convolution layer to extract basic feature representations, before being passed to a series of Residual-in-Residual Dense Blocks (RRDB) [ 24 , 25 ]. The RRDB block is a core component of Real-ESRGAN that combines residual learning and dense connections to effectively capture high-level texture details and maintain training stability in very deep networks. After going through all RRDB blocks, global features are recombined through skip connections and processed in an up sampling stage using a pixel-shuffle mechanism, which efficiently increases the spatial resolution of the image. The final convolution stage produces a high-resolution output image with sharper and more natural visual details, making the Real-ESRGAN architecture very effective for improving the quality of real-world images that have complex degradation. 3.3.3. Efficient Sub-Pixel Convolutional Network (ESPCN) Is a convolutional neural network architecture specifically developed for SR image processing. Unlike conventional approaches that first interpolate low-resolution images before being processed by the network, ESPCN performs the learning and reconstruction process entirely in the low-resolution domain. This approach allows for increased computational efficiency while producing more optimal image recovery quality [ 26 ]. Figure 6 presents the architecture of ESPCN. The Efficient Sub-Pixel Convolutional Network (ESPCN) is a deep learning-based super-resolution (SR) architecture designed to achieve high computational efficiency by performing most convolution operations in the low-resolution (LR) feature space. Unlike conventional SR methods that use bicubic interpolation before feature extraction, ESPCN learns a direct mapping from LR images to high-resolution (HR) outputs through several convolution layers followed by a sub-pixel convolution layer. In this architecture, the LR input image is first processed through multiple convolution layers to extract feature representations efficiently. The resulting feature maps are then passed to a sub-pixel convolution layer that increases spatial resolution by rearranging feature channels through a periodic shuffling operation. For an upscaling factor \(\:r\) , the final convolution layer produces \(\:{r}^{2}\) , feature maps that are reorganized into an HR image. This approach reduces computational complexity and inference time while maintaining competitive reconstruction quality, making ESPCN suitable for real-time and resource-constrained applications such as embedded vision systems and low-power agricultural monitoring platforms. 3.4. Deployment-Oriented System Architecture The proposed system architecture is designed to support real-time SR inference in edge environments with limited computing and power. This architecture is specifically applied to deployment sessions, with the goal of evaluating how the trained SR model behaves when run under real-world operational conditions. Figure 7 presents the system architecture used during the deployment phase to evaluate the behavior of the SR model in an edge environment with limited computing resources and power. The architecture is designed as a distributed vision pipeline, where image acquisition, data transmission, and SR inference are separated to improve efficiency and scalability. The primary inference node runs on an NVIDIA Jetson Orin Nano, which serves as the edge computing platform for executing the super-resolution model. This device supports both CPU and GPU processing, enabling comparative analysis of different inference modes within the same system architecture so that performance differences reflect model characteristics and computing configurations rather than system design. In this system, images are captured every 30 seconds using a low-power ESP32-CAM over a localhost network. The edge device retrieves the images and immediately performs SR inference in real time. Each image is processed individually without batch storage, meaning that every acquisition directly triggers the inference process. Consequently, system latency is determined mainly by the SR model’s processing time rather than scheduling or data queues. This design separates image acquisition frequency from processing response time, where real-time performance refers to the ability of the edge device to generate SR output shortly after image capture. Such an approach is suitable for edge-based monitoring scenarios in which visual data is collected periodically but still requires immediate processing once available. 3.5. Training Model Process Before the model is deployed, it undergoes training and testing. In this study, training and testing were conducted using a previously acquired lettuce plant growth dataset. All models were trained and tested using the same computing environment, the cloud-based Google Colab platform. The process was performed using the PyTorch framework, with CUDA acceleration on an NVIDIA A100 GPU. Previously, we attempted to run the training process on a CPU, but encountered a crash that prevented training. The system used had 83.5 GB of RAM, 40 GB of GPU memory, and 112.6 GB of storage space. The use of a uniform training environment aimed to ensure experimental consistency and enable fair and reproducible performance across the evaluated models. 3.5.1. Dataset Preparation and Pre-Processing Prior to training and evaluation, several dataset preparation and preprocessing steps were performed to improve data quality and ensure accurate image reconstruction. First, histogram normalization was applied to align the pixel intensity distributions between low-resolution (LR) and high-resolution (HR) images. This step is important because evaluation metrics such as PSNR and SSIM are sensitive to differences in intensity and color distribution. Without normalization, metric values may decrease due to lighting or contrast variations rather than the actual SR reconstruction quality [ 27 ]. The training process then used a patch-based approach, where LR–HR image pairs were randomly extracted as 128 × 128 pixel patches. This method improves training efficiency, reduces memory requirements, and increases data diversity, enabling the model to better learn local details and fine textures in lettuce images. For performance evaluation, two separate test sets captured under day and night lighting conditions were used. These datasets were fully separated from the training data to ensure that the evaluation reflects the model’s ability to generalize to real-world conditions, particularly in low-resolution camera-based crop monitoring systems. 3.5.2. Evaluation Matrix The performance of the reconstruction process is quantitatively evaluated using three key metrics: Peak Signal-to-Noise Ratio (PSNR), which measures the ratio of signal strength to noise in decibels (dB), which can be calculated using Eq. 1. $$\:PSNR=10\:\bullet\:\:{log}_{10}\left(\frac{{MAX}_{I}^{2}}{MSE}\right)\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(1\right)$$ Structural Similarity Index (SSIM)Measuring the similarity of structure, luminance, and contrast which is more correlated with human visual perception [ 28 ]. SSIM is very important for assessing the integrity of leaf vein structure, in Eq. 2. $$\:SSIM\left(x,y\right)=\:\frac{\left({2\mu\:}_{x}{\mu\:}_{y}+{C}_{1}\right)\left({2\sigma\:}_{xy}{C}_{2}\right)}{\left({{\mu\:}_{x}}^{2}+{{\mu\:}_{y}}^{2}+{C}_{1}\right)\left({{\sigma\:}_{x}}^{2}+{{\sigma\:}_{y}}^{2}+{C}_{2}\right)}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(2\right)$$ Inference Time: Measures the average time a model takes to process a single input image on the target device, measured in seconds. Measures the total power consumption on the edge computing device as each model processes a digital image. 4. Result and Discussion 4.1. Training Model Evaluation In the training process for each model using epoch configurations of 100, 300 and 500, the aim is to determine the impact of training depth on the processing quality of each SR model. 4.1.1. PSNR Training In PSNR evaluation, a higher PSNR value indicates better image reconstruction quality and lower distortion compared to the reference image. Therefore, the higher the PSNR value, the closer the reconstruction is to the original image [ 29 ]. Table I. Average PSNR value each models Epoch EDSR Real-ESRGAN ESPCN 100 27.96 27.84 26.21 300 28.52 28.59 26.36 500 28.82 28.96 26.47 The training performance of the three super-resolution models is evaluated using the Peak Signal-to-Noise Ratio (PSNR) metric across different training epochs. As illustrated in Fig. 8 , the PSNR trends of the models show a gradual improvement as the number of training epochs increases. During the first 100 epochs, both EDSR and Real-ESRGAN demonstrate a rapid increase in PSNR values, indicating that the models quickly learn the basic reconstruction patterns from the training data. In contrast, ESPCN exhibits a lower PSNR trend and stabilizes earlier, suggesting that its lightweight architecture limits its ability to capture more complex image details compared to the deeper models. The quantitative comparison of the average PSNR values is summarized in Table I. At 100 epochs, EDSR achieves a slightly higher PSNR value (27.96 dB) compared to Real-ESRGAN (27.84 dB), while ESPCN records a noticeably lower value of 26.21 dB. As training continues to 300 epochs, both EDSR and Real-ESRGAN show consistent improvements, reaching 28.52 dB and 28.59 dB respectively. Meanwhile, ESPCN only increases marginally to 26.36 dB, indicating limited reconstruction capability relative to the other models. Further improvements are observed at 500 epochs, where Real-ESRGAN achieves the highest PSNR value of 28.96 dB, slightly outperforming EDSR which reaches 28.82 dB. ESPCN again shows only a small increase to 26.47 dB. These results, as depicted in Fig. 8 and summarized in Table I, indicate that deeper super-resolution architectures such as EDSR and Real-ESRGAN are more effective in reconstructing higher-quality images compared to the lightweight ESPCN model. However, ESPCN still demonstrates stable performance with lower computational complexity, which may be beneficial for resource-constrained edge environments. 4.1.2. SSIM Training In addition to PSNR, the Structural Similarity Index (SSIM) is used to evaluate the quality of reconstructed images during training. Unlike pixel-based metrics, SSIM measures image similarity by considering structure, luminance, and contrast, making it more consistent with human visual perception. It evaluates image degradation based on structural changes rather than absolute pixel differences, making it suitable for assessing reconstruction performance in super-resolution tasks [ 26 ]. In this study, SSIM values are monitored across training epochs to analyze model learning behavior and convergence stability. Higher SSIM values indicate better preservation of structural details in the reconstructed images, such as leaf edges and vein patterns that are important for crop monitoring applications. Table II. Average SSIM value each models Epoch EDSR Real-ESRGAN ESPCN 100 0.803 0.800 0.748 300 0.813 0.816 0.753 500 0.817 0.823 0.757 The training performance of the super-resolution models is further evaluated using the Structural Similarity Index (SSIM). As shown in Fig. 9 , all models exhibit increasing SSIM values as training progresses, indicating improved structural reconstruction capability. At 100 epochs, EDSR and Real-ESRGAN rapidly achieve SSIM values above 0.80, while ESPCN shows lower performance. The quantitative results in Table II confirm this trend, where EDSR and Real-ESRGAN obtain similar SSIM values at early epochs (0.803 and 0.800), while ESPCN records a lower value of 0.748. As training continues, Real-ESRGAN slightly outperforms EDSR, reaching the highest SSIM of 0.823 at 500 epochs, followed by EDSR (0.817) and ESPCN (0.757). The curves in Fig. 9 also indicate that EDSR and Real-ESRGAN approach convergence after approximately 350–400 epochs, whereas ESPCN shows a flatter learning curve due to its lightweight architecture. These results suggest that deeper architectures are more effective in preserving structural details such as leaf edges and vein patterns that are essential for crop monitoring. Furthermore, a consistent relationship between PSNR and SSIM is observed, where models achieving higher PSNR values also obtain higher SSIM scores, as presented in Fig. 8 , Fig. 9 , Table I, and Table II. This indicates that Real-ESRGAN and EDSR provide superior reconstruction quality for low-resolution plant images, while ESPCN remains advantageous for resource-constrained edge environments due to its computational efficiency [ 26 ]. Overall, the experimental results indicate that deeper super-resolution architectures provide better reconstruction performance for low-resolution plant images. Both EDSR and Real-ESRGAN consistently achieve higher PSNR and SSIM values compared to ESPCN, demonstrating their stronger ability to recover fine structural details from degraded inputs. Real-ESRGAN shows a slight advantage at higher training epochs, suggesting that the adversarial learning mechanism contributes to improved perceptual and structural reconstruction quality. Meanwhile, ESPCN exhibits lower reconstruction accuracy but maintains stable training behavior and significantly lower architectural complexity. These findings highlight a practical trade-off between reconstruction quality and computational efficiency. Therefore, while Real-ESRGAN and EDSR are more suitable for applications requiring higher visual fidelity, ESPCN may still be advantageous for deployment in resource-constrained edge environments where lightweight models are preferred. 4.2. Deployment-Oriented Super-Resolution Models 4.2.1. Digital Image Processing Results This testing is a follow-up to the previous model training process. The model, trained with various epoch variations, is saved in .pth format embedded in edge computing and given low-resolution image input (800 x 600 px) using an ESP32-CAM camera. This stage aims to identify the most feasible SR model for implementation on edge computing devices with limited resources. Therefore, on the edge computing side, testing is conducted on two types of hardware environments: CPU and GPU, and in digital image processing there are day and night digital image conditions. From the results of the model training carried out previously, based on the PSNR value evaluation matrix, at epoch 500 each model achieved its best performance, therefore at the digital image processing stage the model used was the model that had been trained with epoch 500. Figure 10 presents a qualitative comparison of the super-resolution reconstruction results generated by EDSR, Real-ESRGAN, and ESPCN from a low-resolution SVGA input image (800 × 600 pixels), which is upscaled to 1600 × 1200 pixels. The enlarged regions highlight the models’ ability to recover fine structural details of the lettuce plant, particularly around the leaf edges and vein structures. EDSR produces a relatively stable reconstruction with smooth textures and consistent structural continuity, indicating its strong capability in minimizing pixel-level reconstruction errors. However, the smoothing effect slightly reduces the visibility of very fine textures in the magnified region. Real-ESRGAN demonstrates sharper edge reconstruction and enhanced high-frequency details, especially along leaf boundaries and small plant structures. This improvement is attributed to its adversarial learning mechanism, which encourages the model to generate perceptually sharper features that resemble natural high-resolution textures. In contrast, ESPCN shows more noticeable blurring and loss of structural clarity in the zoomed area, indicating a reduced ability to reconstruct subtle spatial information from the low-resolution input. These visual differences are further supported by the quantitative PSNR evaluation results. Real-ESRGAN achieves the highest PSNR value of 29.08 dB, indicating the closest reconstruction similarity to the reference high-resolution image. EDSR follows with 28.13 dB, demonstrating competitive reconstruction performance but with slightly lower fidelity in recovering fine textures. Meanwhile, ESPCN records the lowest PSNR value of 27.07 dB, reflecting greater reconstruction error and reduced preservation of structural details. The combined qualitative and quantitative analysis indicates that Real-ESRGAN provides the most effective balance between perceptual sharpness and reconstruction accuracy, making it more suitable for enhancing low-resolution plant images in practical agricultural monitoring scenarios using low-cost camera sensors. Figure 11 presents the digital image processing results of each SR model under nighttime lighting conditions using model weights trained for 500 epochs. The evaluation results show a significant improvement in deployment performance, particularly for Real-ESRGAN, which achieves the highest deployment PSNR of 29.08 dB. Notably, the deployment PSNR values of Real-ESRGAN and ESPCN exceed their corresponding training PSNR values, indicating that under low visual complexity conditions, the models are able to suppress pixel-wise errors more effectively. EDSR maintains stable performance with only a minimal PSNR reduction, confirming its robustness to illumination variations. These results indicate that the effectiveness of long-term training is highly dependent on the operational context, reinforcing the importance of deployment-aware evaluation in edge-based SR system design. 4.2.2. Evaluation of Computational Load of SR Model on Edge Computing This section discusses and evaluates the computational load characteristics of each SR model when deployed on edge devices in real-world scenarios. The evaluation focuses on computational resource utilization and inference processing time, represented by usage (%) and processing time (ms). This analysis aims to identify differences in computational behavior among the models and to assess their implications for real-time processing feasibility in resource-constrained edge environments. Within this context, computational load is considered not merely as a system-level attribute, but as a critical factor in understanding the trade-offs between model complexity, processing responsiveness, and deployment readiness. Table III. Computational load and real-time feasibility per single inference on the edge computing Epoch Model GPU Usage (%) CPU Usage (%) Time (ms) Feasibility 500 ESPCN 2.6 71.9 3 Real-time 500 Real-ESRGAN 97 98.3 119 Conditional / Not feasible 500 EDSR 9.8 52.2 25 Near real-time Table III presents the computational load and real-time feasibility of three super-resolution models when executed on the edge computing platform using the trained models at epoch 500. The results indicate significant differences in hardware utilization and inference latency among the evaluated models. ESPCN demonstrates the most lightweight computational requirement, with GPU usage of only 2.6% and CPU usage of 71.9%, while achieving the fastest inference time of 3 ms. This performance allows ESPCN to operate comfortably within real-time constraints, making it highly suitable for edge-based image enhancement scenarios. In contrast, Real-ESRGAN shows extremely high hardware utilization, reaching 97% GPU usage and 98.3% CPU usage, with an inference time of 119 ms. Although the model produces high-quality reconstruction, the heavy computational demand significantly limits its practicality for continuous real-time deployment on resource-constrained edge devices, therefore categorized as conditionally feasible or not feasible for real-time processing. Meanwhile, EDSR provides a balanced trade-off between reconstruction capability and computational efficiency, with moderate GPU and CPU usage of 9.8% and 52.2%, respectively, and an inference time of 25 ms. This latency places EDSR in the near real-time category, indicating that it can still be applied in edge-based applications with minor latency tolerance. Overall, the analysis highlights that lightweight architectures such as ESPCN are more appropriate for strict real-time edge computing environments, whereas deeper models like Real-ESRGAN impose substantial computational overhead that reduces their deployment feasibility. 4.2.3. Power Consumption of SR Inference on Edge Computing The evaluation was conducted by recording the temporal power profile in milliwatts (mW) across the entire edge computing environment, during a single inference cycle, encompassing the pre-processing idle state, the image inference phase, and the post-processing idle state. This approach allows direct observation of power spikes, processing duration, and recovery patterns back to idle, thus capturing realistic power consumption behavior under real-world deployment conditions. Table IV. Overall edge computing power consumption during SR image processing Model Hardware Avg. Edge Power (W) ESPCN GPU 4.8 CPU 5.6 EDSR GPU 5.1 CPU 11.5 Real-ESRGAN GPU 7.6 CPU 14.2 Figure 12 and Table IV presents the power consumption of the edge computing system during super-resolution (SR) image processing using three different models. The results show clear differences in energy demand depending on the model complexity and the hardware resource utilized. ESPCN demonstrates the lowest power consumption among the evaluated models, with an average of 4.8 W when utilizing the GPU and 5.6 W when processed on the CPU. This indicates that ESPCN is highly energy-efficient and well-suited for deployment on resource-constrained edge devices that require continuous operation. In comparison, EDSR requires moderately higher energy, consuming approximately 5.1 W on the GPU and 11.5 W on the CPU. The increase in CPU power usage reflects the higher computational complexity of the model, although the GPU still maintains relatively efficient energy utilization. Meanwhile, Real-ESRGAN exhibits the highest power consumption, reaching 7.6 W on the GPU and 14.2 W on the CPU. This substantial energy requirement is consistent with the model’s deeper architecture and heavier computational workload, which significantly increases hardware utilization during inference. Overall, the analysis indicates that lightweight models such as ESPCN offer superior energy efficiency for edge computing environments, while more complex architectures like Real-ESRGAN impose greater power demands that may limit their suitability for long-term or energy-sensitive edge deployments. 5. Conclusion This study successfully enhanced and reconstructed digital images of lettuce plants directly using low-resolution camera input. Three deep learning-based super resolution models—EDSR, Real-ESRGAN, and ESPCN—were applied to double the digital image resolution from 800 × 600 pixels (SVGA) to 1600 × 1200 pixels (UXGA) using a lettuce plant growth dataset acquired directly with an ESP32-CAM and implemented on an NVIDIA Jetson Orin Nano-based edge device. The results show that all models are capable of improving image resolution, but reconstruction quality, computational overhead, and power consumption are significantly affected by the complexity of the model architecture and the hardware configuration used. Real-ESRGAN produced the highest visual quality at the expense of the highest computational overhead and power consumption. EDSR offered a balance between quality and resource efficiency, while ESPCN demonstrated the highest computational and energy efficiency, despite limitations in image detail. The implementation evaluation confirmed the gap between training and inference performance on edge devices and demonstrated that GPU acceleration is a key factor in achieving real-time and energy-efficient super resolution inference. Overall, this study confirms that SR model selection must be tailored to the application objectives and resource constraints of edge devices, as no single model is optimal for all scenarios. Further research could focus on developing SR model selection mechanisms that adapt to system conditions and edge device resource constraints, for example through dynamic model selection or adaptive inference strategies. Furthermore, exploring further optimizations such as the use of mixed-precision inference, model quantization, and integrating SR with advanced vision tasks, such as crop detection or segmentation, could potentially improve overall system efficiency without compromising image analysis quality. Declarations Declaration of Generative AI and AI-Assisted Technologies in the Manuscript Preparation Process During the preparation of this manuscript, ChatGPT (OpenAI) was utilized to assist in improving the English language quality, clarity of scientific expression, and grammatical structure. The authors carefully reviewed and edited all generated content and take full responsibility for the final version of the manuscript. Acknowledgements The authors express their sincere appreciation to Universitas Trisakti for providing internal research funding that supported the completion of this study. The authors also gratefully acknowledge the institutional facilities and resources that enabled the implementation of the research activities and experimental work. Author Contribution Mhd. Idham Khalif contributed to conceptualization, methodology, data collection, analysis, and writing of the original draft. Tjhawa Endang Djuana contributed to data analysis and manuscript writing. Richard Antonius Rambung contributed to data collection, image processing analysis, and manuscript writing. Achmad Nadratan Al Janna contributed to machine learning model training and testing. Listyo Edi Prabwo contributed to manuscript editing, formatting, grammar checking, and language revision. Tirta Akdi Toma Mesoya Hulu contributed to manuscript editing, formatting, grammar checking, and language revision. All authors reviewed and approved the final version of the manuscript. Conflict of Interest The authors declare that they have no conflict of interest. Data Availability The dataset used in this study consists of real-world images captured using an ESP32-CAM within an agricultural monitoring environment. Due to data usage and management considerations, the dataset is not publicly accessible. Nevertheless, it may be made available for academic and research purposes upon reasonable request to the corresponding author. References M. Raj and M. Prahadeeswaran. Revolutionizing agriculture: a review of smart farming technologies for a sustainable future. Discover Applied Sciences, 7:937, 2025. https://doi.org/10.1007/s42452-025-07561-6 T. Chen and H. Yin. Camera-based plant growth monitoring for automated plant cultivation with controlled environment agriculture. Smart Agricultural Technology, 8:100449, 2024. https://doi.org/10.1016/j.atech.2024.100449 Y. Szoke and G. Shani. Tracking plant growth using image sequence analysis. Agriculture Communications, 3(4):100110, 2025. https://doi.org/10.1016/j.agrcom.2025.100110 E. Mavridou, E. Vrochidou, G. A. Papakostas, T. Pachidis, and V. G. Kaburlasos. Machine vision systems in precision agriculture for crop farming. Journal of Imaging, 5(12):89, 2019. https://doi.org/10.3390/jimaging5120089 M. Barjaktarovic, M. Santoni, and L. Bruzzone. Design and verification of a low-cost multispectral camera for precision agriculture application. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:6945–6957, 2024. https://doi.org/10.1109/JSTARS.2024.3377104 M. B. Stuart, M. Davies, M. J. Hobbs, T. D. Pering, A. J. S. McGonigle, and J. R. Willmott. High-resolution hyperspectral imaging using low-cost components: Application within environmental monitoring scenarios. Sensors, 22(12):4652, 2022. https://doi.org/10.3390/s22124652 Khalif, I., & Mardian, R. (2025). Comparative Analysis of YOLOv3, MobileNet-SSD, and EfficientDet for Real-Time Person Detection in Low-Resolution Images. ITEGAM-JETIA, 11(55), 293-303. https://doi.org/10.5935/jetia.v11i55.2714 W. Xi, Z. J. Z. Abidin, C. Peng, and T. E. Nyamasvisva. A review of deep learning-based image super-resolution reconstruction methods. Journal of Computing and Electronic Information Management, 17(2):5–11, 2025. https://doi.org/10.54097/phfrck02 Y. K. Ooi and H. Ibrahim. Deep learning algorithms for single image super-resolution: A systematic review. Electronics, 10(7):867, 2021. https://doi.org/10.3390/electronics10070867 H. Su, Y. Li, Y. Xu, X. Fu, and S. Liu. A review of deep-learning-based super-resolution:From methods to applications. Pattern Recognition, 157:110935, 2025a. https://doi.org/10.1016/j.patcog.2024.110935 Yi Rong, Mingbin Jia, Yufei Zhan, and Luoyu Zhou. Sr-rdfan-log: Arbitrary-scale logging image super-resolution reconstruction based on residual dense feature aggregation. Geoenergy Science and Engineering, 240:213042, 2024. ISSN 2949-8910. https://doi.org/10.1016/j.geoen.2024.213042 Yuan Chen, Sitian Li, Hongwei Ma, Peichao Li, Guangming Zhang, Jingjie Guo, and Ming Dong. Super-resolution reconstruction of the holographically reconstructed image based on improved esrgan. Optics Communications, 596:132451, 2025. ISSN 0030-4018. https://doi.org/10.1016/j.optcom.2025.132451 Ağalday, M.F.; Çinar, A. Improvement of a Subpixel Convolutional Neural Network for a Super-Resolution Image. Appl. Sci. 2025, 15, 2459. https://doi.org/10.3390/app15052459 W. Zhang, M. Awais, S. M. Z. A. Naqvi, Y. Xiong, L. Li, Y. Zang, S. Ahmed, J. Wu, H. Zhang, M. I. Abdulraheem, V. Raghavan, J. Ping, and J. Hu. Real-time remote corn growth monitoring system using plant wearable fiber bragg grating sensor. Computers and Electronics in Agriculture, 227:109538, 2024. https://doi.org/10.1016/j.compag.2024.109538 T. Chen and H. Yin. Camera-based plant growth monitoring for automated plant cultivation with controlled environment agriculture. Smart Agricultural Technology, 8:100449, 2024. https://doi.org/10.1016/j.atech.2024.100449 P. Vigneault, J. Lafond-Lapalme, A. Deshaies, K. Khun, S. de la Sablonni`ere, M. Filion, L. Longchamps, and B. Mimee. An integrated data-driven approach to monitor and estimate plant-scale growth using uav. ISPRS Open Journal of Photogrammetry and Remote Sensing, 11:100052, 2024. https://doi.org/10.1016/j.ophoto.2023.100052 B.-K. Xie, S.-B. Liu, and L. Li. Large-scale microscope with improved resolution using SRGAN. Optics & Laser Technology, 179:111291, 2024. https://doi.org/10.1016/j.optlastec.2024.111291 D. Varga, Z. Szab´o, and P. J. Szab´o. Super-resolution enhancement of x-ray microscopic images of solder joints. NDT & E International, 154:103382, 2025. https://doi.org/10.1016/j.ndteint.2025.103382 Z. Ma, B. Bijeljic, G. Wen, et al. Super-resolution imaging of multiphase fluid distributions in porous media using deep learning. Transport in Porous Media, 152:85, 2025. https://doi.org/10.1007/s11242-025-02210-3 Hu Su, Ying Li, Yifan Xu, Xiang Fu, and Song Liu. A review of deep-learning-based super-resolution: From methods to applications. Pattern Recognition, 157:110935, 2025. https://doi.org/10.1016/j.patcog.2024.110935 B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 136–144, 2017. https://doi.org/10.1109/CVPRW.2017.151 S. M. A. Bashir, Y. Wang, M. Khan, and Y. Niu. A comprehensive review of deep learningbased single image super-resolution. PeerJ Computer Science, 7:e621, 2021. https://doi.org/10.7717/peerj-cs.621 K. Zeng, H. Zheng, C. Cai, Y. Yang, K. Zhang, and Z. Chen. Simultaneous single- and multicontrast super-resolution for brain mri images based on a convolutional neural network. Computers in Biology and Medicine, 99:133–141, 2018. https://doi.org/10.1016/j.compbiomed.2018.06.010 X. Wang, L. Xie, C. Dong, and Y. Shan. Real-esrgan: Training real-world blind superresolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1905–1914. IEEE, 2021. https://doi.org/10.1109/ICCVW54120.2021.00217 J. Guerreiro, P. Tom´as, N. Garcia, and H. Aidos. Super-resolution of magnetic resonance images using generative adversarial networks. Computerized Medical Imaging and Graphics, 108:102280, 2023. https://doi.org/10.1016/j.compmedimag.2023.102280 Zhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600–612, 2004. https://doi.org/10.1109/TIP.2003.819861 M. Dohmen, M. A. Klemens, I. M. Baltruschat, et al. Similarity and quality metrics for mr image-to-image translation. Scientific Reports, 15:3853, 2025. https://doi.org/10.1038/s41598-025-87358-0 Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., 2004. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612. https://doi.org/10.1109/TIP.2003.819861 Horé, A., & Ziou, D. (2010). Image quality metrics: PSNR vs. SSIM. Proceedings of the 20th International Conference on Pattern Recognition, 2366–2369. https://doi.org/10.1109/ICPR.2010.579 Additional Declarations No competing interests reported. Supplementary Files 4GraphicalAbstract.pdf 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-9305065","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":618388570,"identity":"4c786271-5bea-4f8b-94e7-9ede3a8f02fa","order_by":0,"name":"Mhd. Idham Khalif","email":"data:image/png;base64,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","orcid":"","institution":"Universitas Trisakti","correspondingAuthor":true,"prefix":"","firstName":"Mhd.","middleName":"Idham","lastName":"Khalif","suffix":""},{"id":618388574,"identity":"2c7beeea-d9ab-4b34-a1b8-3291fdad7711","order_by":1,"name":"Tjhwa Endang Djuana","email":"","orcid":"","institution":"Universitas Trisakti","correspondingAuthor":false,"prefix":"","firstName":"Tjhwa","middleName":"Endang","lastName":"Djuana","suffix":""},{"id":618388575,"identity":"c91264ab-6603-416d-b0fc-782b9b19608f","order_by":2,"name":"Richard Antonius Rambung","email":"","orcid":"","institution":"Universitas Trisakti","correspondingAuthor":false,"prefix":"","firstName":"Richard","middleName":"Antonius","lastName":"Rambung","suffix":""},{"id":618388576,"identity":"f05c2a4a-478f-4160-bd44-44fc10192250","order_by":3,"name":"Achmad Nadratan Al Janna","email":"","orcid":"","institution":"Universitas Trisakti","correspondingAuthor":false,"prefix":"","firstName":"Achmad","middleName":"Nadratan Al","lastName":"Janna","suffix":""},{"id":618388578,"identity":"db3bb17f-2333-4748-afcf-bd8c1493b044","order_by":4,"name":"Listyo Edi Prabowo","email":"","orcid":"","institution":"Universitas Indonesia","correspondingAuthor":false,"prefix":"","firstName":"Listyo","middleName":"Edi","lastName":"Prabowo","suffix":""},{"id":618388581,"identity":"1fc69fed-6236-4005-a022-bbf56ae14f64","order_by":5,"name":"Tirta Akdi Toma Mesoya Hulu","email":"","orcid":"","institution":"The University of Queensland","correspondingAuthor":false,"prefix":"","firstName":"Tirta","middleName":"Akdi Toma Mesoya","lastName":"Hulu","suffix":""}],"badges":[],"createdAt":"2026-04-02 15:38:34","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9305065/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9305065/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106877856,"identity":"b2b6f713-14ae-4ef7-ba00-18417d6fc838","added_by":"auto","created_at":"2026-04-14 10:42:36","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":4847809,"visible":true,"origin":"","legend":"\u003cp\u003ePlacement and configuration of ESP32-CAM in image data capture\u003c/p\u003e","description":"","filename":"Fig1Configurationimagedatacapture.png","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/f5d44db91822db06b2494e2b.png"},{"id":106877751,"identity":"e176987f-07cf-450a-a025-04b844db011f","added_by":"auto","created_at":"2026-04-14 10:42:30","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":99070,"visible":true,"origin":"","legend":"\u003cp\u003eCorrupted UXGA resolution results\u003c/p\u003e","description":"","filename":"Fig2CorruptedUXGAresolutionresults.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/146a388fc118c829ee8821f8.jpg"},{"id":106877828,"identity":"857143f3-af41-4fc1-9898-8227bae852dc","added_by":"auto","created_at":"2026-04-14 10:42:33","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1301736,"visible":true,"origin":"","legend":"\u003cp\u003eSample of dataset\u003c/p\u003e","description":"","filename":"Fig3Sampledataset.png","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/bca0ba07ecfd7864369d078d.png"},{"id":106877763,"identity":"aa81e28b-187d-45c1-b13e-5d2054c021c2","added_by":"auto","created_at":"2026-04-14 10:42:31","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":111780,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of EDSR\u003c/p\u003e","description":"","filename":"Fig4ArchofEDSR.png","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/a091099b61810423f8d06739.png"},{"id":106877752,"identity":"285132ee-af80-4aa4-a0aa-17911e899ef8","added_by":"auto","created_at":"2026-04-14 10:42:30","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":218296,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of Real-ESRGAN\u003c/p\u003e","description":"","filename":"Fig5ArchofRealESRGAN1.png","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/87655fb868a71ba2aebae5f9.png"},{"id":106877854,"identity":"cafc353f-0e64-4472-8717-43b18646fae9","added_by":"auto","created_at":"2026-04-14 10:42:36","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1884514,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of ESPCN\u003c/p\u003e","description":"","filename":"Fig6ArchofESPCNArch.png","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/013e31c3e550bf6855a6e748.png"},{"id":106877915,"identity":"5c854ae4-7eb1-479f-a828-87421acd2a95","added_by":"auto","created_at":"2026-04-14 10:42:47","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":804465,"visible":true,"origin":"","legend":"\u003cp\u003eSystem architecture on deployment\u003c/p\u003e","description":"","filename":"Fig7SystemArch.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/accea7f1e20d10d3abd28fc9.jpg"},{"id":106877846,"identity":"2ebef889-45b8-41f7-a7e7-532bc0e8858b","added_by":"auto","created_at":"2026-04-14 10:42:34","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":1200752,"visible":true,"origin":"","legend":"\u003cp\u003eGraph comparison of training PSNR\u003c/p\u003e","description":"","filename":"Fig8PSNRGraph.png","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/87719ae1d2ecf7cc7feffbf9.png"},{"id":106877760,"identity":"b778271f-62f1-4638-9c09-ffd0ecc0ad94","added_by":"auto","created_at":"2026-04-14 10:42:31","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":1331143,"visible":true,"origin":"","legend":"\u003cp\u003eGraph comparison of training SSIM\u003c/p\u003e","description":"","filename":"Fig9SSIMGraph.png","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/44e998e907cb6906db1ab522.png"},{"id":106877765,"identity":"c1d4b9e1-02cc-49de-914c-d759d9108c52","added_by":"auto","created_at":"2026-04-14 10:42:31","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":1122315,"visible":true,"origin":"","legend":"\u003cp\u003eImage processing results for each model with 500 epochs of day conditions\u003c/p\u003e","description":"","filename":"Fig10ResultDaytime.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/d53d2a315e0920a46bd58d9a.jpg"},{"id":106877762,"identity":"5f605c42-6e1a-4ba9-ad67-86594eb8535f","added_by":"auto","created_at":"2026-04-14 10:42:31","extension":"jpg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":1063189,"visible":true,"origin":"","legend":"\u003cp\u003eImage processing results for each model with 500 epochs of night conditions\u003c/p\u003e","description":"","filename":"Fig11ResultNighttime.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/a0a4bfe15dc09cf9f83824ba.jpg"},{"id":106877852,"identity":"9046c64d-a78a-449a-8a99-84a1a626c33b","added_by":"auto","created_at":"2026-04-14 10:42:36","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":649374,"visible":true,"origin":"","legend":"\u003cp\u003ePower Consumption of SR Inference on Edge Computing\u003c/p\u003e","description":"","filename":"Fig12PowerConsumption.png","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/4996bf831c6cc79ffab3837b.png"},{"id":106994369,"identity":"ac85bbbb-673e-4e76-82b0-5852447a1c0d","added_by":"auto","created_at":"2026-04-15 15:08:04","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":21368450,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/93b66649-0bfd-49e3-9198-bc0017bfb468.pdf"},{"id":106877845,"identity":"ccad4241-0849-48f5-a631-e310f0f80f6f","added_by":"auto","created_at":"2026-04-14 10:42:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":2338498,"visible":true,"origin":"","legend":"","description":"","filename":"4GraphicalAbstract.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9305065/v1/14166a56e83cacf8f5bd9d6d.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Bridging the Gap Between Low-Cost Cameras and High-Fidelity Monitoring: Deployment of Super-Resolution Models for Real-World Lettuce Farming","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003ePlant growth monitoring is a fundamental component of modern agricultural systems, enabling continuous observation of plant development, health assessment, and early detection of productivity-threatening disturbances [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Advances in digital technology have accelerated the adoption of smart agriculture, where Artificial Intelligence (AI) plays a key role in improving efficiency and crop yields.\u003c/p\u003e \u003cp\u003eCamera-based image monitoring enables continuous and non-invasive data acquisition without manual intervention [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. However, high-quality imaging systems often involve high costs and energy consumption, limiting their scalability in resource-constrained environments [\u003cspan additionalcitationids=\"CR5\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe ESP32-CAM has emerged as a low-cost alternative due to its compact size, wireless connectivity, and suitability for edge computing applications [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Nevertheless, its limited image resolution can negatively affect downstream tasks such as plant detection and analysis.\u003c/p\u003e \u003cp\u003eDeep learning-based super-resolution (SR) techniques offer a promising solution by reconstructing high-resolution images from low-resolution inputs without increasing hardware cost. Although SR models have demonstrated strong performance [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], most are trained on synthetic datasets that do not reflect real-world degradation [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis study applies to edge computing, three SR architectures\u0026mdash;EDSR [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], Real-ESRGAN [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], and ESPCN [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], using real lettuce image dataset captured by ESP32-CAM, so that it can be evaluated on the feasibility of implementation in edge-based vision system, so that it can be applied to low-cost visual precision lettuce crop monitoring system with optimal resource usage.\u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Vision-Based Plant Growth Monitoring\u003c/h2\u003e \u003cp\u003eVarious studies have applied computer vision-based technology to monitor plant growth in real-time. One early study utilized a Raspberry Pi-based camera to non-invasively monitor corn cob growth over a long period of time, and showed that the image-based system was able to capture plant growth dynamics on an ongoing basis [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Other research has developed a plant growth monitoring system using a web camera connected to a computer, where the image results are processed to model plant structure in three dimensions (3D), thus enabling more in-depth analysis of plant morphology [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Apart from that, the use of drones equipped with cameras has also been widely applied to monitor plant growth from an aerial perspective, especially in large-scale agricultural areas [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAlthough this camera-based approach has proven effective in monitoring plant growth, most of the systems developed still rely on camera devices with high specifications, both in terms of image resolution and processing capabilities. This dependency generally results in increased initial investment costs and system energy consumption. These conditions pose significant challenges to the implementation of crop monitoring systems that are sustainable, cost-effective and widely applicable, especially in agricultural environments with limited resources.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Super-Resolution (SR) for Image Enhancement\u003c/h2\u003e \u003cp\u003eSR techniques have been widely applied as a method for improving image quality (image enhancement) in various application domains. In the field of biomedical imaging, the Real-ESRGAN method is used to increase the resolution of tissue microscopic images, where low resolution images with a wide field of view (FOV) are successfully reconstructed into high resolution images without losing FOV, while expanding the depth of field (DOF) to a certain extent [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Another approach uses Efficient Sub-Pixel Convolutional Network (ESPCN) to increase the resolution of X-ray microscopic images, with evaluation results showing a significant increase in image quality and an average Peak Signal-to-Noise Ratio (PSNR) value reaching 40 dB [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. In addition, Enhanced Deep Super Resolution (EDSR) has also been applied to three-dimensional (3D) X-ray images of porous media, resulting in image reconstruction with sharper details and clearer structures, and proving the effectiveness of EDSR in complex scientific imaging domains [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Research Gap and Contribution\u003c/h2\u003e \u003cp\u003ePrevious studies show that although computer vision and deep learning-based super-resolution (SR) techniques have been widely applied in various imaging fields, research evaluating SR model performance on plant images captured directly using low-cost cameras under real conditions remains limited. Most studies rely on synthetic datasets or high-specification imaging devices, which do not fully represent the real-world degradation characteristics of cameras such as ESP32-CAM. Moreover, comparative analyses that consider not only reconstructed image quality but also training efficiency and the potential implementation in cost- and energy-efficient crop monitoring systems\u0026mdash;particularly for lettuce\u0026mdash;are still scarce. Therefore, this study contributes by providing a comprehensive evaluation of three deep learning-based SR models\u0026mdash;EDSR, Real-ESRGAN, and ESPCN\u0026mdash;using lettuce image datasets captured directly from ESP32-CAM. The evaluation considers image quality, training dynamics, and deployment potential on edge computing, aiming to support the selection of appropriate SR models for low-cost camera-based plant monitoring systems in smart precision agriculture.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Methodology","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Data Acquisition Setup\u003c/h2\u003e \u003cp\u003eIn this study, the image data acquisition process was carried out using an ESP32-CAM device as the main image capture system, thus representing the real conditions of continuous lettuce plant growth monitoring over time. The selection of the ESP32-CAM was based on its characteristics as a low-cost camera widely used in Internet of Things (IoT) applications and smart agriculture systems. Despite having limitations in sensor resolution, dynamic range, and image quality stability, this device reflects the real challenges faced in implementing low-cost camera-based crop monitoring systems. Therefore, the resulting dataset has the characteristics of low-resolution images with realistic visual degradation, making it very relevant to evaluate the effectiveness of deep learning-based SR methods in improving crop image quality without the need for expensive hardware.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the use of the ESP32-CAM as the image acquisition device in this study. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e(a) shows the placement of the ESP32-CAM on the growing medium to continuously document mustard plant growth during the observation period, while Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e(b) presents the front view of the camera setup. The camera was positioned at a fixed angle and distance from the plant to ensure consistent data acquisition. Images were captured every 30 minutes to record visual changes in plant appearance due to growth and environmental variations during both day and night. In each session, the ESP32-CAM sequentially captured two images with a one-second interval: an 800 \u0026times; 600 pixel (SVGA) image as the low-resolution (LR) dataset and a 1600 \u0026times; 1200 pixel (UXGA) image as the high-resolution (HR) dataset.\u003c/p\u003e \u003cp\u003eSVGA resolution was selected as the LR representation because it provides a balance between visual detail and device limitations. Compared with VGA or lower resolutions, SVGA retains sufficient morphological information of plants, making it suitable for developing and evaluating SR techniques on low-cost cameras. Although the ESP32-CAM supports UXGA resolution, it is not always stable in practice, and some captured images may be corrupted compared to those at SVGA resolution, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Dataset Construction\u003c/h2\u003e \u003cp\u003eThe dataset used in this study consists of a total of 2,505 low-resolution (LR) and high-resolution (HR) image pairs. All image pairs were acquired through the same data acquisition process and under identical capture conditions, so each LR image has a spatially and temporally corresponding HR image pair. This approach, similar to paired datasets captured in the real world, has been shown to be important in SR research, as DL models rely on accurate mapping between LR and HR images to learn features and detect degradation, and because model performance often degrades when trained solely on synthetic images that do not reflect real-world conditions [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis approach ensures that visual differences between LR and HR images are caused solely by differences in sensor resolution, not by external factors such as lighting changes, object movement, or variations in camera angle. Therefore, the resulting dataset is wellsuited for training and evaluating deep learning-based SR methods, as it allows for an objective and representative assessment of model performance against real-world conditions in low-cost crop monitoring systems.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows an example of a digital image dataset of lettuce plant growth used in this study. In general, the dataset was collected under two main lighting conditions: daytime, with adequate natural light intensity, and night time, with minimal ambient lighting and the aid of a flash. These varying lighting conditions are intended to represent real-world image acquisition conditions.\u003c/p\u003e \u003cp\u003eIn the dataset used, low-resolution (LR) and high-resolution (HR) image pairs served as the basis for evaluating the performance of a SR-based deep learning model. This evaluation aimed to assess the model\u0026rsquo;s ability to improve the quality of low-resolution images to approximate high-resolution reference images. The use of a relatively large paired dataset is expected to yield a more representative and reliable evaluation of model performance in real-world scenarios. Each LR\u0026ndash;HR image pair was obtained from the same image capture session at two consecutive resolutions. This approach aims to ensure that the image pairs originate from identical acquisition conditions, ensuring that differences in image quality are solely due to differences in sensor resolution, rather than external factors such as changes in lighting or object movement.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Super-Resolution (SR) Methods\u003c/h2\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e3.3.1. Enhanced Deep Super-Resolution (EDSR)\u003c/h2\u003e \u003cp\u003eIs a deep learning-based SR technique designed to improve the quality of low-resolution images into high-resolution images by utilizing Convolutional Neural Networks (CNNs) which are known for their good accuracy. EDSR eliminates non-essential components in conventional SR architectures, such as batch normalization, thus improving feature representation capabilities and producing image reconstructions with sharper details and higher accuracy [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the Enhanced Deep Super-Resolution (EDSR) architecture [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], for 2x SR, consists of several key components designed to effectively reconstruct high-resolution (HR) images from low-resolution (LR) images. EDSR is trained using an end-to-end deep learning approach that allows feature learning to be performed directly from low-resolution images to high-resolution ones.\u003c/p\u003e \u003cp\u003eThis approach results in superior reconstruction performance, making EDSR known as one of the best-performing CNN-based architectures in SR tasks [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. The process begins with an initial convolution layer that extracts basic features from the input image. These features are then processed through a series of residual blocks without batch normalization (ResBlock), as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, to enhance feature representation while preserving the range of pixel intensity values. Each residual block employs a skip connection to maintain training stability and retain important information from previous layers.\u003c/p\u003e \u003cp\u003eAfter passing through the residual blocks, the extracted features are combined with the initial features through a global residual connection before entering the upsampling stage. At this stage, EDSR applies a sub-pixel convolution (pixel shuffle) mechanism, illustrated in the Up sample section of Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, to efficiently increase spatial resolution. This architectural design enables EDSR to produce sharper image reconstructions with 2\u0026times; enhancement, making it suitable for improving low-resolution lettuce images captured using low-cost cameras such as ESP32-CAM.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e3.3.2. Real-Enhanced Super-Resolution Generative Adversial Network (Real-ESRGAN)\u003c/h2\u003e \u003cp\u003eIs one of the methods in SR known as perceptual quality, to improve the visual quality of images that experience complex degradation in the real world through a generative model. This model is trained with synthetic data that simulates real degradation so that it is able to improve visual details and reduce artifacts better than previous methods on real datasets [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e is the Real-ESRGAN architecture, the SR process begins with a pixel-unshuffle stage on low-resolution input images, especially for magnification scales of \u0026times;2 and \u0026times;1. This operation functions to reverse the pixel-shuffle mechanism by reducing the spatial resolution of the image while increasing the number of feature channels, so that the computational load can be reduced without losing important information. The pixel-unshuffled features are then processed through an initial convolution layer to extract basic feature representations, before being passed to a series of Residual-in-Residual Dense Blocks (RRDB) [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe RRDB block is a core component of Real-ESRGAN that combines residual learning and dense connections to effectively capture high-level texture details and maintain training stability in very deep networks. After going through all RRDB blocks, global features are recombined through skip connections and processed in an up sampling stage using a pixel-shuffle mechanism, which efficiently increases the spatial resolution of the image. The final convolution stage produces a high-resolution output image with sharper and more natural visual details, making the Real-ESRGAN architecture very effective for improving the quality of real-world images that have complex degradation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.3.3. Efficient Sub-Pixel Convolutional Network (ESPCN)\u003c/h2\u003e \u003cp\u003eIs a convolutional neural network architecture specifically developed for SR image processing. Unlike conventional approaches that first interpolate low-resolution images before being processed by the network, ESPCN performs the learning and reconstruction process entirely in the low-resolution domain. This approach allows for increased computational efficiency while producing more optimal image recovery quality [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e presents the architecture of ESPCN. The Efficient Sub-Pixel Convolutional Network (ESPCN) is a deep learning-based super-resolution (SR) architecture designed to achieve high computational efficiency by performing most convolution operations in the low-resolution (LR) feature space. Unlike conventional SR methods that use bicubic interpolation before feature extraction, ESPCN learns a direct mapping from LR images to high-resolution (HR) outputs through several convolution layers followed by a sub-pixel convolution layer.\u003c/p\u003e \u003cp\u003eIn this architecture, the LR input image is first processed through multiple convolution layers to extract feature representations efficiently. The resulting feature maps are then passed to a sub-pixel convolution layer that increases spatial resolution by rearranging feature channels through a periodic shuffling operation. For an upscaling factor \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:r\\)\u003c/span\u003e\u003c/span\u003e, the final convolution layer produces \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{r}^{2}\\)\u003c/span\u003e\u003c/span\u003e, feature maps that are reorganized into an HR image. This approach reduces computational complexity and inference time while maintaining competitive reconstruction quality, making ESPCN suitable for real-time and resource-constrained applications such as embedded vision systems and low-power agricultural monitoring platforms.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.4. Deployment-Oriented System Architecture\u003c/h2\u003e \u003cp\u003eThe proposed system architecture is designed to support real-time SR inference in edge environments with limited computing and power. This architecture is specifically applied to deployment sessions, with the goal of evaluating how the trained SR model behaves when run under real-world operational conditions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e presents the system architecture used during the deployment phase to evaluate the behavior of the SR model in an edge environment with limited computing resources and power. The architecture is designed as a distributed vision pipeline, where image acquisition, data transmission, and SR inference are separated to improve efficiency and scalability.\u003c/p\u003e \u003cp\u003eThe primary inference node runs on an NVIDIA Jetson Orin Nano, which serves as the edge computing platform for executing the super-resolution model. This device supports both CPU and GPU processing, enabling comparative analysis of different inference modes within the same system architecture so that performance differences reflect model characteristics and computing configurations rather than system design.\u003c/p\u003e \u003cp\u003eIn this system, images are captured every 30 seconds using a low-power ESP32-CAM over a localhost network. The edge device retrieves the images and immediately performs SR inference in real time. Each image is processed individually without batch storage, meaning that every acquisition directly triggers the inference process. Consequently, system latency is determined mainly by the SR model\u0026rsquo;s processing time rather than scheduling or data queues.\u003c/p\u003e \u003cp\u003eThis design separates image acquisition frequency from processing response time, where real-time performance refers to the ability of the edge device to generate SR output shortly after image capture. Such an approach is suitable for edge-based monitoring scenarios in which visual data is collected periodically but still requires immediate processing once available.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.5. Training Model Process\u003c/h2\u003e \u003cp\u003eBefore the model is deployed, it undergoes training and testing. In this study, training and testing were conducted using a previously acquired lettuce plant growth dataset. All models were trained and tested using the same computing environment, the cloud-based Google Colab platform. The process was performed using the PyTorch framework, with CUDA acceleration on an NVIDIA A100 GPU. Previously, we attempted to run the training process on a CPU, but encountered a crash that prevented training. The system used had 83.5 GB of RAM, 40 GB of GPU memory, and 112.6 GB of storage space. The use of a uniform training environment aimed to ensure experimental consistency and enable fair and reproducible performance across the evaluated models.\u003c/p\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e3.5.1. Dataset Preparation and Pre-Processing\u003c/h2\u003e \u003cp\u003ePrior to training and evaluation, several dataset preparation and preprocessing steps were performed to improve data quality and ensure accurate image reconstruction. First, histogram normalization was applied to align the pixel intensity distributions between low-resolution (LR) and high-resolution (HR) images. This step is important because evaluation metrics such as PSNR and SSIM are sensitive to differences in intensity and color distribution. Without normalization, metric values may decrease due to lighting or contrast variations rather than the actual SR reconstruction quality [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe training process then used a patch-based approach, where LR\u0026ndash;HR image pairs were randomly extracted as 128 \u0026times; 128 pixel patches. This method improves training efficiency, reduces memory requirements, and increases data diversity, enabling the model to better learn local details and fine textures in lettuce images.\u003c/p\u003e \u003cp\u003eFor performance evaluation, two separate test sets captured under day and night lighting conditions were used. These datasets were fully separated from the training data to ensure that the evaluation reflects the model\u0026rsquo;s ability to generalize to real-world conditions, particularly in low-resolution camera-based crop monitoring systems.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e3.5.2. Evaluation Matrix\u003c/h2\u003e \u003cp\u003eThe performance of the reconstruction process is quantitatively evaluated using three key metrics:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003ePeak Signal-to-Noise Ratio (PSNR), which measures the ratio of signal strength to noise in decibels (dB), which can be calculated using Eq.\u0026nbsp;1.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv id=\"Equa\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:PSNR=10\\:\\bullet\\:\\:{log}_{10}\\left(\\frac{{MAX}_{I}^{2}}{MSE}\\right)\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(1\\right)$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eStructural Similarity Index (SSIM)Measuring the similarity of structure, luminance, and contrast which is more correlated with human visual perception [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. SSIM is very important for assessing the integrity of leaf vein structure, in Eq.\u0026nbsp;2.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv id=\"Equb\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:SSIM\\left(x,y\\right)=\\:\\frac{\\left({2\\mu\\:}_{x}{\\mu\\:}_{y}+{C}_{1}\\right)\\left({2\\sigma\\:}_{xy}{C}_{2}\\right)}{\\left({{\\mu\\:}_{x}}^{2}+{{\\mu\\:}_{y}}^{2}+{C}_{1}\\right)\\left({{\\sigma\\:}_{x}}^{2}+{{\\sigma\\:}_{y}}^{2}+{C}_{2}\\right)}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(2\\right)$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eInference Time: Measures the average time a model takes to process a single input image on the target device, measured in seconds.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eMeasures the total power consumption on the edge computing device as each model processes a digital image.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"4. Result and Discussion","content":"\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Training Model Evaluation\u003c/h2\u003e \u003cp\u003eIn the training process for each model using epoch configurations of 100, 300 and 500, the aim is to determine the impact of training depth on the processing quality of each SR model.\u003c/p\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e4.1.1. PSNR Training\u003c/h2\u003e \u003cp\u003eIn PSNR evaluation, a higher PSNR value indicates better image reconstruction quality and lower distortion compared to the reference image. Therefore, the higher the PSNR value, the closer the reconstruction is to the original image [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eTable I.\u003c/b\u003e Average PSNR value each models\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"4\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEpoch\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEDSR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReal-ESRGAN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eESPCN\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003e100\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e27.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e27.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e26.21\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003e300\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e28.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e28.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e26.36\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003e500\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e28.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e28.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e26.47\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\u003eThe training performance of the three super-resolution models is evaluated using the Peak Signal-to-Noise Ratio (PSNR) metric across different training epochs. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, the PSNR trends of the models show a gradual improvement as the number of training epochs increases. During the first 100 epochs, both EDSR and Real-ESRGAN demonstrate a rapid increase in PSNR values, indicating that the models quickly learn the basic reconstruction patterns from the training data. In contrast, ESPCN exhibits a lower PSNR trend and stabilizes earlier, suggesting that its lightweight architecture limits its ability to capture more complex image details compared to the deeper models.\u003c/p\u003e \u003cp\u003eThe quantitative comparison of the average PSNR values is summarized in Table I. At 100 epochs, EDSR achieves a slightly higher PSNR value (27.96 dB) compared to Real-ESRGAN (27.84 dB), while ESPCN records a noticeably lower value of 26.21 dB. As training continues to 300 epochs, both EDSR and Real-ESRGAN show consistent improvements, reaching 28.52 dB and 28.59 dB respectively. Meanwhile, ESPCN only increases marginally to 26.36 dB, indicating limited reconstruction capability relative to the other models.\u003c/p\u003e \u003cp\u003eFurther improvements are observed at 500 epochs, where Real-ESRGAN achieves the highest PSNR value of 28.96 dB, slightly outperforming EDSR which reaches 28.82 dB. ESPCN again shows only a small increase to 26.47 dB. These results, as depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e and summarized in Table I, indicate that deeper super-resolution architectures such as EDSR and Real-ESRGAN are more effective in reconstructing higher-quality images compared to the lightweight ESPCN model. However, ESPCN still demonstrates stable performance with lower computational complexity, which may be beneficial for resource-constrained edge environments.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e4.1.2. SSIM Training\u003c/h2\u003e \u003cp\u003eIn addition to PSNR, the Structural Similarity Index (SSIM) is used to evaluate the quality of reconstructed images during training. Unlike pixel-based metrics, SSIM measures image similarity by considering structure, luminance, and contrast, making it more consistent with human visual perception. It evaluates image degradation based on structural changes rather than absolute pixel differences, making it suitable for assessing reconstruction performance in super-resolution tasks [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. In this study, SSIM values are monitored across training epochs to analyze model learning behavior and convergence stability. Higher SSIM values indicate better preservation of structural details in the reconstructed images, such as leaf edges and vein patterns that are important for crop monitoring applications.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eTable II.\u003c/b\u003e Average SSIM value each models\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"4\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEpoch\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEDSR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReal-ESRGAN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eESPCN\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003e100\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.803\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.800\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.748\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003e300\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.813\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.816\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.753\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003e500\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.817\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.823\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.757\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\u003eThe training performance of the super-resolution models is further evaluated using the Structural Similarity Index (SSIM). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e, all models exhibit increasing SSIM values as training progresses, indicating improved structural reconstruction capability. At 100 epochs, EDSR and Real-ESRGAN rapidly achieve SSIM values above 0.80, while ESPCN shows lower performance. The quantitative results in Table II confirm this trend, where EDSR and Real-ESRGAN obtain similar SSIM values at early epochs (0.803 and 0.800), while ESPCN records a lower value of 0.748. As training continues, Real-ESRGAN slightly outperforms EDSR, reaching the highest SSIM of 0.823 at 500 epochs, followed by EDSR (0.817) and ESPCN (0.757).\u003c/p\u003e \u003cp\u003eThe curves in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e also indicate that EDSR and Real-ESRGAN approach convergence after approximately 350\u0026ndash;400 epochs, whereas ESPCN shows a flatter learning curve due to its lightweight architecture. These results suggest that deeper architectures are more effective in preserving structural details such as leaf edges and vein patterns that are essential for crop monitoring. Furthermore, a consistent relationship between PSNR and SSIM is observed, where models achieving higher PSNR values also obtain higher SSIM scores, as presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e, Table I, and Table II. This indicates that Real-ESRGAN and EDSR provide superior reconstruction quality for low-resolution plant images, while ESPCN remains advantageous for resource-constrained edge environments due to its computational efficiency [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOverall, the experimental results indicate that deeper super-resolution architectures provide better reconstruction performance for low-resolution plant images. Both EDSR and Real-ESRGAN consistently achieve higher PSNR and SSIM values compared to ESPCN, demonstrating their stronger ability to recover fine structural details from degraded inputs. Real-ESRGAN shows a slight advantage at higher training epochs, suggesting that the adversarial learning mechanism contributes to improved perceptual and structural reconstruction quality. Meanwhile, ESPCN exhibits lower reconstruction accuracy but maintains stable training behavior and significantly lower architectural complexity. These findings highlight a practical trade-off between reconstruction quality and computational efficiency. Therefore, while Real-ESRGAN and EDSR are more suitable for applications requiring higher visual fidelity, ESPCN may still be advantageous for deployment in resource-constrained edge environments where lightweight models are preferred.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Deployment-Oriented Super-Resolution Models\u003c/h2\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e4.2.1. Digital Image Processing Results\u003c/h2\u003e \u003cp\u003eThis testing is a follow-up to the previous model training process. The model, trained with various epoch variations, is saved in .pth format embedded in edge computing and given low-resolution image input (800 x 600 px) using an ESP32-CAM camera. This stage aims to identify the most feasible SR model for implementation on edge computing devices with limited resources. Therefore, on the edge computing side, testing is conducted on two types of hardware environments: CPU and GPU, and in digital image processing there are day and night digital image conditions.\u003c/p\u003e \u003cp\u003eFrom the results of the model training carried out previously, based on the PSNR value evaluation matrix, at epoch 500 each model achieved its best performance, therefore at the digital image processing stage the model used was the model that had been trained with epoch 500.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e presents a qualitative comparison of the super-resolution reconstruction results generated by EDSR, Real-ESRGAN, and ESPCN from a low-resolution SVGA input image (800 \u0026times; 600 pixels), which is upscaled to 1600 \u0026times; 1200 pixels. The enlarged regions highlight the models\u0026rsquo; ability to recover fine structural details of the lettuce plant, particularly around the leaf edges and vein structures. EDSR produces a relatively stable reconstruction with smooth textures and consistent structural continuity, indicating its strong capability in minimizing pixel-level reconstruction errors. However, the smoothing effect slightly reduces the visibility of very fine textures in the magnified region. Real-ESRGAN demonstrates sharper edge reconstruction and enhanced high-frequency details, especially along leaf boundaries and small plant structures. This improvement is attributed to its adversarial learning mechanism, which encourages the model to generate perceptually sharper features that resemble natural high-resolution textures. In contrast, ESPCN shows more noticeable blurring and loss of structural clarity in the zoomed area, indicating a reduced ability to reconstruct subtle spatial information from the low-resolution input.\u003c/p\u003e \u003cp\u003eThese visual differences are further supported by the quantitative PSNR evaluation results. Real-ESRGAN achieves the highest PSNR value of 29.08 dB, indicating the closest reconstruction similarity to the reference high-resolution image. EDSR follows with 28.13 dB, demonstrating competitive reconstruction performance but with slightly lower fidelity in recovering fine textures. Meanwhile, ESPCN records the lowest PSNR value of 27.07 dB, reflecting greater reconstruction error and reduced preservation of structural details. The combined qualitative and quantitative analysis indicates that Real-ESRGAN provides the most effective balance between perceptual sharpness and reconstruction accuracy, making it more suitable for enhancing low-resolution plant images in practical agricultural monitoring scenarios using low-cost camera sensors.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e presents the digital image processing results of each SR model under nighttime lighting conditions using model weights trained for 500 epochs. The evaluation results show a significant improvement in deployment performance, particularly for Real-ESRGAN, which achieves the highest deployment PSNR of 29.08 dB. Notably, the deployment PSNR values of Real-ESRGAN and ESPCN exceed their corresponding training PSNR values, indicating that under low visual complexity conditions, the models are able to suppress pixel-wise errors more effectively. EDSR maintains stable performance with only a minimal PSNR reduction, confirming its robustness to illumination variations. These results indicate that the effectiveness of long-term training is highly dependent on the operational context, reinforcing the importance of deployment-aware evaluation in edge-based SR system design.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e4.2.2. Evaluation of Computational Load of SR Model on Edge Computing\u003c/h2\u003e \u003cp\u003eThis section discusses and evaluates the computational load characteristics of each SR model when deployed on edge devices in real-world scenarios. The evaluation focuses on computational resource utilization and inference processing time, represented by usage (%) and processing time (ms). This analysis aims to identify differences in computational behavior among the models and to assess their implications for real-time processing feasibility in resource-constrained edge environments. Within this context, computational load is considered not merely as a system-level attribute, but as a critical factor in understanding the trade-offs between model complexity, processing responsiveness, and deployment readiness.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTable III.\u003c/b\u003e Computational load and real-time feasibility per single inference on the edge computing\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\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=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEpoch\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGPU Usage (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCPU Usage (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTime (ms)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eFeasibility\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e500\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eESPCN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e71.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eReal-time\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e500\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal-ESRGAN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e98.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e119\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eConditional / Not feasible\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e500\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEDSR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e52.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNear real-time\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\u003eTable III presents the computational load and real-time feasibility of three super-resolution models when executed on the edge computing platform using the trained models at epoch 500. The results indicate significant differences in hardware utilization and inference latency among the evaluated models. ESPCN demonstrates the most lightweight computational requirement, with GPU usage of only 2.6% and CPU usage of 71.9%, while achieving the fastest inference time of 3 ms. This performance allows ESPCN to operate comfortably within real-time constraints, making it highly suitable for edge-based image enhancement scenarios. In contrast, Real-ESRGAN shows extremely high hardware utilization, reaching 97% GPU usage and 98.3% CPU usage, with an inference time of 119 ms. Although the model produces high-quality reconstruction, the heavy computational demand significantly limits its practicality for continuous real-time deployment on resource-constrained edge devices, therefore categorized as conditionally feasible or not feasible for real-time processing. Meanwhile, EDSR provides a balanced trade-off between reconstruction capability and computational efficiency, with moderate GPU and CPU usage of 9.8% and 52.2%, respectively, and an inference time of 25 ms. This latency places EDSR in the near real-time category, indicating that it can still be applied in edge-based applications with minor latency tolerance. Overall, the analysis highlights that lightweight architectures such as ESPCN are more appropriate for strict real-time edge computing environments, whereas deeper models like Real-ESRGAN impose substantial computational overhead that reduces their deployment feasibility.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e4.2.3. Power Consumption of SR Inference on Edge Computing\u003c/h2\u003e \u003cp\u003eThe evaluation was conducted by recording the temporal power profile in milliwatts (mW) across the entire edge computing environment, during a single inference cycle, encompassing the pre-processing idle state, the image inference phase, and the post-processing idle state. This approach allows direct observation of power spikes, processing duration, and recovery patterns back to idle, thus capturing realistic power consumption behavior under real-world deployment conditions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eTable IV.\u003c/b\u003e Overall edge computing power consumption during SR image processing\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"304\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eHardware\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e\u003cstrong\u003eAvg. Edge Power (W)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003eESPCN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eGPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eCPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003eEDSR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eGPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e5.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eCPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e11.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\"\u003e\n \u003cp\u003eReal-ESRGAN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003eGPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e7.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd\u003e\n \u003cp\u003eCPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd\u003e\n \u003cp\u003e14.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eFigure 12 and Table IV presents the power consumption of the edge computing system during super-resolution (SR) image processing using three different models. The results show clear differences in energy demand depending on the model complexity and the hardware resource utilized. ESPCN demonstrates the lowest power consumption among the evaluated models, with an average of 4.8 W when utilizing the GPU and 5.6 W when processed on the CPU. This indicates that ESPCN is highly energy-efficient and well-suited for deployment on resource-constrained edge devices that require continuous operation. In comparison, EDSR requires moderately higher energy, consuming approximately 5.1 W on the GPU and 11.5 W on the CPU. The increase in CPU power usage reflects the higher computational complexity of the model, although the GPU still maintains relatively efficient energy utilization. Meanwhile, Real-ESRGAN exhibits the highest power consumption, reaching 7.6 W on the GPU and 14.2 W on the CPU. This substantial energy requirement is consistent with the model\u0026rsquo;s deeper architecture and heavier computational workload, which significantly increases hardware utilization during inference. Overall, the analysis indicates that lightweight models such as ESPCN offer superior energy efficiency for edge computing environments, while more complex architectures like Real-ESRGAN impose greater power demands that may limit their suitability for long-term or energy-sensitive edge deployments.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study successfully enhanced and reconstructed digital images of lettuce plants directly using low-resolution camera input. Three deep learning-based super resolution models\u0026mdash;EDSR, Real-ESRGAN, and ESPCN\u0026mdash;were applied to double the digital image resolution from 800 \u0026times; 600 pixels (SVGA) to 1600 \u0026times; 1200 pixels (UXGA) using a lettuce plant growth dataset acquired directly with an ESP32-CAM and implemented on an NVIDIA Jetson Orin Nano-based edge device. The results show that all models are capable of improving image resolution, but reconstruction quality, computational overhead, and power consumption are significantly affected by the complexity of the model architecture and the hardware configuration used.\u003c/p\u003e \u003cp\u003eReal-ESRGAN produced the highest visual quality at the expense of the highest computational overhead and power consumption. EDSR offered a balance between quality and resource efficiency, while ESPCN demonstrated the highest computational and energy efficiency, despite limitations in image detail. The implementation evaluation confirmed the gap between training and inference performance on edge devices and demonstrated that GPU acceleration is a key factor in achieving real-time and energy-efficient super resolution inference. Overall, this study confirms that SR model selection must be tailored to the application objectives and resource constraints of edge devices, as no single model is optimal for all scenarios.\u003c/p\u003e \u003cp\u003eFurther research could focus on developing SR model selection mechanisms that adapt to system conditions and edge device resource constraints, for example through dynamic model selection or adaptive inference strategies. Furthermore, exploring further optimizations such as the use of mixed-precision inference, model quantization, and integrating SR with advanced vision tasks, such as crop detection or segmentation, could potentially improve overall system efficiency without compromising image analysis quality.\u003c/p\u003e "},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eDeclaration of Generative AI and AI-Assisted Technologies in the Manuscript Preparation Process\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this manuscript, ChatGPT (OpenAI) was utilized to assist in improving the English language quality, clarity of scientific expression, and grammatical structure. The authors carefully reviewed and edited all generated content and take full responsibility for the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors express their sincere appreciation to Universitas Trisakti for providing internal research funding that supported the completion of this study. The authors also gratefully acknowledge the institutional facilities and resources that enabled the implementation of the research activities and experimental work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMhd. Idham Khalif contributed to conceptualization, methodology, data collection, analysis, and writing of the original draft. Tjhawa Endang Djuana contributed to data analysis and manuscript writing. Richard Antonius Rambung contributed to data collection, image processing analysis, and manuscript writing. Achmad Nadratan Al Janna contributed to machine learning model training and testing. Listyo Edi Prabwo contributed to manuscript editing, formatting, grammar checking, and language revision. Tirta Akdi Toma Mesoya Hulu contributed to manuscript editing, formatting, grammar checking, and language revision. All authors reviewed and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of Interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset used in this study consists of real-world images captured using an ESP32-CAM within an agricultural monitoring environment. Due to data usage and management considerations, the dataset is not publicly accessible. Nevertheless, it may be made available for academic and research purposes upon reasonable request to the corresponding author.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eM. Raj and M. Prahadeeswaran. Revolutionizing agriculture: a review of smart farming technologies for a sustainable future. Discover Applied Sciences, 7:937, 2025. https://doi.org/10.1007/s42452-025-07561-6\u003c/li\u003e\n\u003cli\u003eT. Chen and H. Yin. Camera-based plant growth monitoring for automated plant cultivation with controlled environment agriculture. Smart Agricultural Technology, 8:100449, 2024. https://doi.org/10.1016/j.atech.2024.100449\u003c/li\u003e\n\u003cli\u003eY. Szoke and G. Shani. Tracking plant growth using image sequence analysis. Agriculture Communications, 3(4):100110, 2025. https://doi.org/10.1016/j.agrcom.2025.100110\u003c/li\u003e\n\u003cli\u003eE. Mavridou, E. Vrochidou, G. A. Papakostas, T. Pachidis, and V. G. Kaburlasos. Machine vision systems in precision agriculture for crop farming. Journal of Imaging, 5(12):89, 2019. https://doi.org/10.3390/jimaging5120089\u003c/li\u003e\n\u003cli\u003eM. Barjaktarovic, M. Santoni, and L. Bruzzone. Design and verification of a low-cost multispectral camera for precision agriculture application. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:6945\u0026ndash;6957, 2024. https://doi.org/10.1109/JSTARS.2024.3377104\u003c/li\u003e\n\u003cli\u003eM. B. Stuart, M. Davies, M. J. Hobbs, T. D. Pering, A. J. S. McGonigle, and J. R. Willmott. High-resolution hyperspectral imaging using low-cost components: Application within environmental monitoring scenarios. Sensors, 22(12):4652, 2022. https://doi.org/10.3390/s22124652\u003c/li\u003e\n\u003cli\u003eKhalif, I., \u0026amp; Mardian, R. (2025). Comparative Analysis of YOLOv3, MobileNet-SSD, and EfficientDet for Real-Time Person Detection in Low-Resolution Images. ITEGAM-JETIA, 11(55), 293-303. https://doi.org/10.5935/jetia.v11i55.2714\u003c/li\u003e\n\u003cli\u003eW. Xi, Z. J. Z. Abidin, C. Peng, and T. E. Nyamasvisva. A review of deep learning-based image super-resolution reconstruction methods. Journal of Computing and Electronic Information Management, 17(2):5\u0026ndash;11, 2025. https://doi.org/10.54097/phfrck02\u003c/li\u003e\n\u003cli\u003eY. K. Ooi and H. Ibrahim. Deep learning algorithms for single image super-resolution: A systematic review. Electronics, 10(7):867, 2021. https://doi.org/10.3390/electronics10070867\u003c/li\u003e\n\u003cli\u003eH. Su, Y. Li, Y. Xu, X. Fu, and S. Liu. A review of deep-learning-based super-resolution:From methods to applications. Pattern Recognition, 157:110935, 2025a. https://doi.org/10.1016/j.patcog.2024.110935\u003c/li\u003e\n\u003cli\u003eYi Rong, Mingbin Jia, Yufei Zhan, and Luoyu Zhou. Sr-rdfan-log: Arbitrary-scale logging image super-resolution reconstruction based on residual dense feature aggregation. Geoenergy Science and Engineering, 240:213042, 2024. ISSN 2949-8910. https://doi.org/10.1016/j.geoen.2024.213042\u003c/li\u003e\n\u003cli\u003eYuan Chen, Sitian Li, Hongwei Ma, Peichao Li, Guangming Zhang, Jingjie Guo, and Ming Dong. Super-resolution reconstruction of the holographically reconstructed image based on improved esrgan. Optics Communications, 596:132451, 2025. ISSN 0030-4018. https://doi.org/10.1016/j.optcom.2025.132451\u003c/li\u003e\n\u003cli\u003eAğalday, M.F.; \u0026Ccedil;inar, A. Improvement of a Subpixel Convolutional Neural Network for a Super-Resolution Image. Appl. Sci. 2025, 15, 2459. https://doi.org/10.3390/app15052459\u003c/li\u003e\n\u003cli\u003eW. Zhang, M. Awais, S. M. Z. A. Naqvi, Y. Xiong, L. Li, Y. Zang, S. Ahmed, J. Wu, H. Zhang, M. I. Abdulraheem, V. Raghavan, J. Ping, and J. Hu. Real-time remote corn growth monitoring system using plant wearable fiber bragg grating sensor. Computers and Electronics in Agriculture, 227:109538, 2024. https://doi.org/10.1016/j.compag.2024.109538\u003c/li\u003e\n\u003cli\u003eT. Chen and H. Yin. Camera-based plant growth monitoring for automated plant cultivation with controlled environment agriculture. Smart Agricultural Technology, 8:100449, 2024. https://doi.org/10.1016/j.atech.2024.100449\u003c/li\u003e\n\u003cli\u003eP. Vigneault, J. Lafond-Lapalme, A. Deshaies, K. Khun, S. de la Sablonni`ere, M. Filion, L. Longchamps, and B. Mimee. An integrated data-driven approach to monitor and estimate plant-scale growth using uav. ISPRS Open Journal of Photogrammetry and Remote Sensing, 11:100052, 2024. https://doi.org/10.1016/j.ophoto.2023.100052\u003c/li\u003e\n\u003cli\u003eB.-K. Xie, S.-B. Liu, and L. Li. Large-scale microscope with improved resolution using SRGAN. Optics \u0026amp; Laser Technology, 179:111291, 2024. https://doi.org/10.1016/j.optlastec.2024.111291\u003c/li\u003e\n\u003cli\u003eD. Varga, Z. Szab\u0026acute;o, and P. J. Szab\u0026acute;o. Super-resolution enhancement of x-ray microscopic images of solder joints. NDT \u0026amp; E International, 154:103382, 2025. https://doi.org/10.1016/j.ndteint.2025.103382\u003c/li\u003e\n\u003cli\u003eZ. Ma, B. Bijeljic, G. Wen, et al. Super-resolution imaging of multiphase fluid distributions in porous media using deep learning. Transport in Porous Media, 152:85, 2025. https://doi.org/10.1007/s11242-025-02210-3\u003c/li\u003e\n\u003cli\u003eHu Su, Ying Li, Yifan Xu, Xiang Fu, and Song Liu. A review of deep-learning-based super-resolution: From methods to applications. Pattern Recognition, 157:110935, 2025. https://doi.org/10.1016/j.patcog.2024.110935\u003c/li\u003e\n\u003cli\u003eB. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 136\u0026ndash;144, 2017. https://doi.org/10.1109/CVPRW.2017.151\u003c/li\u003e\n\u003cli\u003eS. M. A. Bashir, Y. Wang, M. Khan, and Y. Niu. A comprehensive review of deep learningbased single image super-resolution. PeerJ Computer Science, 7:e621, 2021. https://doi.org/10.7717/peerj-cs.621\u003c/li\u003e\n\u003cli\u003eK. Zeng, H. Zheng, C. Cai, Y. Yang, K. Zhang, and Z. Chen. Simultaneous single- and multicontrast super-resolution for brain mri images based on a convolutional neural network. Computers in Biology and Medicine, 99:133\u0026ndash;141, 2018. https://doi.org/10.1016/j.compbiomed.2018.06.010\u003c/li\u003e\n\u003cli\u003eX. Wang, L. Xie, C. Dong, and Y. Shan. Real-esrgan: Training real-world blind superresolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1905\u0026ndash;1914. IEEE, 2021. https://doi.org/10.1109/ICCVW54120.2021.00217\u003c/li\u003e\n\u003cli\u003eJ. Guerreiro, P. Tom\u0026acute;as, N. Garcia, and H. Aidos. Super-resolution of magnetic resonance images using generative adversarial networks. Computerized Medical Imaging and Graphics, 108:102280, 2023. https://doi.org/10.1016/j.compmedimag.2023.102280\u003c/li\u003e\n\u003cli\u003eZhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600\u0026ndash;612, 2004. https://doi.org/10.1109/TIP.2003.819861\u003c/li\u003e\n\u003cli\u003eM. Dohmen, M. A. Klemens, I. M. Baltruschat, et al. Similarity and quality metrics for mr image-to-image translation. Scientific Reports, 15:3853, 2025. https://doi.org/10.1038/s41598-025-87358-0\u003c/li\u003e\n\u003cli\u003eWang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., 2004. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600\u0026ndash;612. https://doi.org/10.1109/TIP.2003.819861\u003c/li\u003e\n\u003cli\u003eHor\u0026eacute;, A., \u0026amp; Ziou, D. (2010). Image quality metrics: PSNR vs. SSIM. Proceedings of the 20th International Conference on Pattern Recognition, 2366\u0026ndash;2369. https://doi.org/10.1109/ICPR.2010.579\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"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":"super-resolution, edge computing, smart agriculture, deep learning, visual monitoring","lastPublishedDoi":"10.21203/rs.3.rs-9305065/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9305065/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCurrently, the application of visual monitoring in smart agriculture is one of the options in the implementation of precision agriculture. Smart agriculture visual monitoring has challenges in terms of the relatively high cost of high-resolution cameras and limited access to resources. One option that can be used is implementing embedded low-resolution cameras so that the cost is also low and improves the quality of image resolution by implementing a deep learning-based Super-Resolution (SR) method. This study applies image enhancement to low-resolution cameras directly using three deep learning-based SR models\u0026mdash;EDSR, Real-ESRGAN, and ESPCN\u0026mdash;for a 2\u0026times; resolution increase from 800 \u0026times; 600 (SVGA) to 1600 \u0026times; 1200 (UXGA) to find out the SR model that is suitable for precision agriculture according to the conditions. Experiments were conducted using a real-world lettuce growth dataset taken by ESP32-CAM as input to the low-resolution camera and implemented on NVIDIA Jetson Orin Nano as edge computing. Performance was assessed in terms of reconstruction quality, computational load, processing latency, and power consumption under CPU and GPU execution. The results show that Real-ESRGAN achieves the highest visual quality at the expense of computational and energy requirements, EDSR offers a good balance, and ESPCN provides the highest efficiency with reduced image detail. These findings highlight the potential for low-cost visual growth monitoring of lettuce plants under limited resource constraints, leading to precision agriculture applications.\u003c/p\u003e","manuscriptTitle":"Bridging the Gap Between Low-Cost Cameras and High-Fidelity Monitoring: Deployment of Super-Resolution Models for Real-World Lettuce Farming","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-14 10:41:35","doi":"10.21203/rs.3.rs-9305065/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":"8f9d630f-cb67-48ed-8460-4c5895a59706","owner":[],"postedDate":"April 14th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-14T10:42:20+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-14 10:41:35","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9305065","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9305065","identity":"rs-9305065","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}