Identification and Classification of Bronze Surface Diseases Based on Improved YOLOv5

Identification and Classification of Bronze Surface Diseases Based on Improved YOLOv5 | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Identification and Classification of Bronze Surface Diseases Based on Improved YOLOv5 Han Wu, Jing Yang, Lijun Pang, Wei Zhou, Yang Lei, Xuegang Liu, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6900599/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 By leveraging the lightweight and high-performance characteristics of EfficientNetV2, this paper enhances the YOLOv5s model to accurately classify disease types in bronze wares while reducing human error. The efficiency of feature extraction across multiple scales is further improved through an optimized convolutional structure and a compound scaling method. The bronze wares housed in the Jingmen Museum serve as a case study to validate the model’s effectiveness. Photographs of the artifacts were captured using a Canon G1X Mark III digital camera and processed into a dataset comprising 1,037 images, labeled with defects such as surface flaws, holes, full-body mineralization, and cracks for training and validation. Results show that the enhanced YOLOv5s_EfficientNetV2 model significantly outperforms the original YOLOv5s model across key metrics including Precision, Recall, [email protected] , and [email protected] :0.95, with respective improvements of 7.0%, 1.7%, 2.2%, and 2.6%. Furthermore, a set of visual interfaces for disease detection in bronze wares has been developed using OpenCV and PyQt5, allowing intuitive visualization of the results. The study also explores the impact of batch size on model performance, revealing that a batch size of 8 offers the best trade-off between gradient estimation accuracy and training stability, thereby enhancing recognition capability. By quantifying the type and location of deterioration, this technology offers valuable data support for conservation strategies, provides robust assistance in the protection and restoration of bronze artifacts, and demonstrates strong potential for broader applications in cultural heritage preservation. Improved YOLOv5s EfficientNetV2 Bronze Artifact Disease Recognition Object Detection Visual interface 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 Figure 13 Figure 14 Figure 15 Figure 16 1. Introduction Bronzes hold profound historical, cultural, and artistic value and represent a significant component of China’s cultural heritage. However, over time and under the influence of environmental factors, various forms of corrosion, both on the surface and within the structure, have developed [ 1 – 4 ] . Common types of deterioration include mutilation, porosity, full-body mineralization, and fractures. These not only compromise the aesthetic appeal of the artifacts but also pose serious risks to their structural integrity [ 2 , 5 ] . Without timely identification and intervention, such diseases can spread, leading to irreversible damage. Traditional manual inspection methods heavily rely on experience and are prone to subjective error, often resulting in suboptimal restoration strategies. Deep learning-based object detection technologies offer promising solutions by automatically identifying disease boundaries, classifying and locating defect regions, and outputting structured data including coordinates and confidence scores. This capability makes such technologies highly valuable for accurate and timely diagnosis of bronze disease. Conventional detection methods largely depend on visual assessment and expert judgment. These approaches are inefficient, subjective, and inadequate for large-scale artifact conservation. In response, modern detection technologies have been gradually introduced. For example, Liu et al. [ 6 ] employed dual-energy X-ray imaging to reconstruct 3D structures and detect internal damage in bronzes. Ramana M. et al. [ 7 ] utilized wavelet transforms and Fourier analysis on SEM images to quantify pitting and cracking in Ni–Al bronze surfaces by estimating fractal dimensions. While effective, these methods require substantial computational resources and are often based on linear classification, limiting their precision in complex or overlapping cases. Recent advancements in computer vision and deep learning have significantly enhanced performance in image recognition and classification, particularly with complex, high-dimensional data [ 8 – 10 ] . These techniques not only boost detection efficiency but also reduce human error, making them well-suited for the preventive preservation of bronzes. Some research efforts have already applied deep learning to bronze disease identification. For instance, Xia [ 11 ] proposed a method using an improved DenseNet architecture combined with Class Activation Mapping (CAM) to achieve weakly supervised disease localization. By integrating SE Blocks and the CBAM attention module, the model effectively extracted and identified disease features. However, this approach still has limitations: weakly supervised localization methods like ES-CAM, while superior to traditional techniques, rely heavily on feature pyramids and complex post-processing steps with high computational costs, making real-time application challenging. Furthermore, the need for extensive architectural modification complicates implementation [ 12 ] . In contrast, YOLOv5—a lightweight object detection model known for its fast inference and high detection accuracy—is better suited for real-world disease detection scenarios. Xie et al. [ 13 ] introduced a feature-enhanced YOLO framework for detecting surface defects, improving FPN design and spatial correlation at multiple scales. Similarly, Zhao et al. [ 14 ] proposed an enhanced YOLOv5 model for metal surface defect detection, integrating a parameter-free spatial attention (PSA) mechanism that improved feature extraction from key regions and boosted defect detection performance by 2.4% over the baseline. Despite these improvements, the original YOLOv5 model still struggles with dense object detection and small-scale features [ 15 ] . In response to these challenges, this study presents an enhanced YOLOv5s model for detecting and classifying surface diseases in bronze wares, aiming to address the limitations of traditional methods in terms of efficiency and objectivity. The model integrates the lightweight and high-performance EfficientNetV2 backbone, which enhances detection of small and dense targets. By incorporating a compound scaling method and optimized convolution structure, the model improves multi-scale feature extraction while reducing computational cost and parameter size [ 16 , 17 ] . To evaluate the model's effectiveness, high-resolution images of bronze artifacts from the Jingmen Museum collection were used to create a dataset. A total of 1,037 images (600×600 pixels) were labeled with disease categories such as flaws, holes, full-body mineralization, and cracks. After data augmentation, the dataset was split into training and test sets in an 8:2 ratio. The performance of the improved model was compared against the baseline YOLOv5s and an attention-enhanced variant by analyzing key metrics—Precision, Recall, and [email protected] . Additionally, the effect of batch size on model convergence and generalization was investigated. To support practical application, a visual detection interface was developed using OpenCV and PyQt5, enabling automatic disease classification and intuitive spatial visualization of detection results. This provides intelligent technical support for the preservation and restoration of bronze wares. Moreover, the proposed methodology can be extended to other domains involving cultural heritage diagnostics, underscoring its broad practical relevance. 2. Modeling methods and steps To address the limitations of traditional manual detection methods for bronze ware diseases, characterized by low efficiency and high error rates, this study focuses on the bronze artifacts housed in the Jingmen Museum and constructs a dedicated dataset for automated analysis. High-resolution surface images of bronze wares were captured using a Canon G1X Mark III digital camera. In accordance with national classification standards for cultural heritage deterioration, images depicting four typical types of disease, whole-body mineralization, cracks, holes, and surface defects, were selected for further analysis. Preprocessing steps, including uniform cropping to a resolution of 600×600 pixels, rotation, and flipping, were carried out using Adobe Photoshop. To alleviate the problem of overfitting caused by a limited sample size, data augmentation techniques such as random cropping and translation were applied, expanding the dataset to 1,037 images. Disease regions were annotated using the LabelImg tool, generating YOLO-format annotation files containing both disease categories and bounding box coordinates. To improve the original YOLOv5s model’s limitations in small object detection and multi-scale feature fusion, this study integrates EfficientNetV2 as a lightweight backbone network. Model development and training were conducted in PyCharm using the PyTorch framework, with a batch size of 8 determined empirically. To provide an intuitive visualization of the detection results, an interactive GUI application was developed by combining OpenCV and PyQt5. OpenCV handles image preprocessing, bounding box drawing, and category labeling, while PyQt5 is used to construct the interface, offering functionalities such as model import, image upload, and real-time detection. The system enables the localization and classification of diseases in input images, with results presented via bounding boxes and confidence scores. In the model validation phase, detection performance was quantitatively evaluated using standard metrics, including Precision, Recall, and mean Average Precision at IoU thresholds of 0.5 (mAP50) and from 0.5 to 0.95 (mAP50:95). The overall model architecture and workflow are illustrated in Fig. 1 . 2.1 Data sets and pre-processing This study uses bronze wares from the Jingmen Museum collection as representative specimens. High-resolution digital photographs of the artifacts were captured using a Canon G1X Mark III camera. To accommodate the complex and varied surface lighting conditions of the bronze wares, adjustable parameters such as aperture, shutter speed, and ISO were carefully configured. Specifically, an aperture (F-number) between F5.6 and F8 was chosen to strike a balance between image sharpness and depth of field. An excessively small aperture can cause diffraction-related image degradation, while a large aperture may result in loss of detail due to a shallow depth of field. The shutter speed was controlled within the range of 1/100 to 1/200 seconds to reduce motion blur caused by handheld shaking. In low-light conditions, the shutter speed could be slightly reduced, provided the camera was stabilized (e.g., with a tripod), and the lowest possible ISO setting was used to minimize image noise. Image preprocessing, including random cropping, flipping, and rotation, was conducted using Adobe Photoshop. All images were exported in JPG format with a standardized resolution of 600×600 pixels, resulting in a total of 1,037 annotated images depicting four distinct types of bronzeware surface defects. The finalized cropped dataset is available in the Supporting Information (Document S1). LabelImg was used to annotate each image by marking the diseased regions and assigning defect categories. Annotations were saved in the YOLO format, containing information such as the bounding box’s center coordinates (x, y) and its width (W) and height (H). A summary of the annotation results for each image is presented in Table S1 –S2. The dataset was randomly divided into training and prediction (validation) sets in a ratio of 8:2. Representative samples of annotated disease regions are illustrated in Fig. 2. 2.2 Relevant evaluation indicators In this study, the performance of the model is evaluated using four key metrics: Precision, Recall, Average Precision (AP), and mean Average Precision (mAP) at different Intersection over Union (IoU) thresholds, specifically from 0.50 to 0.95. Equations (1) and (2) define Precision and Recall, respectively [ 14 , 18 ] . $$\:Precision=\frac{TP}{TP+FP}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(1\right)$$ $$\:Recall=\frac{TP}{TP+FN}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(2\right)$$ where TP refers to the number of instances correctly identified as positive (i.e., correctly detected diseased regions). FP denotes the number of instances incorrectly predicted as positive (i.e., non-diseased areas mistakenly labeled as diseased). FN represents the number of actual positive instances that the model fails to detect (i.e., diseased areas predicted as healthy) [ 15 , 19 ] . Higher Precision indicates that fewer false positives are present in the predictions, while higher Recall reflects the model’s ability to detect the majority of actual positive samples. Equations (3) and (4) define AP and mean Average Precision ( [email protected] ) [ 20 – 22 ] . $$\:AP={\int\:}_{0}^{1}P\left(R\right)dR\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(3\right)$$ $$\:mAP=\frac{{\sum\:}_{i=1}^{n}A{P}_{i}}{n}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(4\right)$$ The AP for a single class is computed as the area under the Precision–Recall (P–R) curve, and mAP is the mean of AP values across all classes, where 𝑛 is the number of categories considered. In this experiment, [email protected] and [email protected] :0.95 are both reported to provide a comprehensive assessment of the model’s detection performance across varying IoU thresholds. 3. YOLOv5 structure and improvements 3.1 Convolutional Network Neural Convolutional Neural Networks (CNNs) are widely used deep learning algorithms particularly well-suited for processing multidimensional data such as images, speech, and natural language. Their key attributes, local connectivity and parameter sharing, enable them to recognize image features both quickly and accurately, making CNNs highly effective across various tasks and application domains. A typical CNN architecture is composed of five fundamental components: the input layer, convolutional layers, pooling layers, fully connected layers, and the output layer [23] . CNNs operate by alternating between convolutional, pooling, and fully connected layers, which allows them to automatically extract high-level, abstract features from input data. The globally shared weights and locally connected structure of CNNs reduce the number of learnable parameters and enhance computational efficiency, making them particularly adept at handling high-dimensional inputs such as images. Among these components, the convolutional layer plays a central role in feature extraction. It processes the input image using a set of learnable filters (convolutional kernels), which perform localized weighted summations followed by the addition of a bias term. Each kernel is designed to detect specific local features such as edges, textures, or color patterns, enabling hierarchical feature learning as the network depth increases. Following the convolutional layer is the pooling layer, which primarily serves to reduce the spatial dimensions of the data while retaining essential feature information. As illustrated in Figure 3, the most commonly used pooling operations are max pooling and average pooling. Max pooling selects the highest value within a defined window, while average pooling computes the mean of values in that region. Pooling not only reduces the computational burden but also enhances the model’s robustness to small shifts, noise, and distortions in the input. Finally, the fully connected (FC) layer appears near the end of the network and is responsible for mapping the extracted features to the target output space, such as classification labels. Each neuron in the FC layer is connected to every neuron in the preceding layer, allowing the model to integrate spatially distributed features and generate the final prediction. 3.2 Improved YOLOv5s network architecture 3.2.1 YOLOv5s YOLOv5s, a lightweight variant within the YOLO (You Only Look Once) series of object detection models, is widely favored for its efficiency, accuracy, and ease of deployment. Its network architecture is primarily composed of three components: the Backbone, Neck, and Head layers [24, 25] . These components work in conjunction to perform feature extraction, feature fusion, and target detection from the input image. The overall structure of the model is illustrated in Figure 4. The Backbone is responsible for receiving and preprocessing the input image. Within the Backbone, the Focus module performs a slicing operation on the input image, followed by convolutional processing to extract features. By reducing the number of channels and increasing the spatial resolution (i.e., width and height) of the feature map, the Focus module preserves more detailed spatial information. [26] . Another critical component of the Backbone is the CSPNet (Cross Stage Partial Network) module, a lightweight structure designed to improve the learning ability of the network while reducing computational cost [20] . Additionally, the SPP (Spatial Pyramid Pooling) module enhances the detection of objects at multiple scales by performing pooling operations at different receptive field sizes [27, 28] . The Neck of YOLOv5s incorporates both FPN (Feature Pyramid Network) and PANet (Path Aggregation Network) structures. FPN generates multi-scale feature maps by combining features across different levels via lateral connections and up-sampling operations [28] . PANet further strengthens this information flow by introducing bottom-up paths with down-sampling operations, thus improving the fusion of low-level and high-level feature representations [29, 30] . Following feature fusion in the Neck, the Head layer processes the resulting feature maps via convolutional operations to produce final detection outputs. These outputs include bounding box coordinates, objectness scores (confidence levels), and class probabilities, enabling precise target localization and classification. To enhance the model's generalization performance during training, YOLOv5s employs a series of data augmentation techniques such as mosaic augmentation and adaptive anchor box computation. Furthermore, the model's lightweight design significantly reduces its computational cost and number of parameters, making it highly suitable for real-time applications on edge devices. 3.2.2 EfficientNetV2 In 2019, Tan and Le et al. [31] proposed EfficientNet, a family of convolutional neural networks that introduced a novel model scaling approach to achieve a better trade-off between accuracy and efficiency. Their study systematically investigated how scaling the depth, width, and resolution of a network influences its performance. Unlike traditional scaling strategies that focus on a single dimension, EfficientNet adopts a composite scaling method that uniformly scales all three dimensions using a fixed set of scaling coefficients. This technique mitigates the limitations associated with one-dimensional scaling approaches, enabling EfficientNet to achieve state-of-the-art results with significantly fewer parameters and FLOPs. However, EfficientNetV1 exhibited certain limitations during training. For instance, training on large-resolution images was found to be slow and inefficient, particularly in the early layers of the network that employed depthwise separable convolutions [32] . While depthwise convolutions reduce parameter count and computational load, they are often not optimized on current hardware accelerators, resulting in suboptimal performance. Moreover, EfficientNetV1's approach of uniformly scaling all network stages was later shown to be inefficient, as different stages contribute unequally to overall model complexity and training dynamics. To address these issues, EfficientNetV2 introduced several architectural improvements, notably the Fused-MBConv block [33] . In contrast to the original MBConv structure—comprising an expansion 1×1 convolution followed by a depthwise 3×3 convolution—the Fused-MBConv simplifies this pipeline by replacing it with a standard 3×3 convolution, enhancing compatibility with modern mobile and server accelerators (as illustrated in Figure 5). Another key innovation in EfficientNetV2 is its non-uniform scaling strategy, which scales different network stages unequally based on their contribution to model efficiency and training time. This adjustment improves both training speed and parameter utilization. The network begins with standard preprocessing operations, such as normalization and resizing, to prepare the input images. Feature extraction is then performed through successive convolutional and pooling layers, with many of the convolutional layers utilizing depthwise separable convolutions to further reduce computational overhead. To ensure consistent performance across input images of varying resolutions, a feature deflation module is introduced. This module applies global average pooling to convert spatial feature maps into fixed-length vectors, which are then passed through a learnable 1×1 convolution (or linear transformation) to align the feature dimensions. To enhance the network’s representational capacity, multi-scale feature fusion strategies, such as feature cascading or aggregation, are applied to combine feature maps from different depths [34] . The final stage of the network is typically a fully connected layer (or a final convolutional layer, depending on the task), which outputs either a probability distribution over class labels (using the softmax activation function for classification tasks) or continuous values such as bounding box coordinates for regression tasks. The training process involves computing a loss function-typically cross-entropy for classification or mean squared error (MSE) for regression-based on the discrepancy between predicted outputs and ground truth labels. Network parameters are then iteratively updated using backpropagation to minimize this loss, thereby optimizing the model’s predictive performance. 3.2.3 YOLOv5s_EfficientNetV2 This study proposes the integration of EfficientNetV2 into the YOLOv5s framework, forming a YOLOv5s_EfficientNetV2 hybrid network to enhance the baseline model’s feature extraction capability. EfficientNetV2, with its lightweight architecture and composite scaling strategy, is designed to balance model size, computational efficiency, and representational power. By optimizing the network structure and parameters, EfficientNetV2 is able to retain strong feature extraction performance while reducing the computational burden—making it highly suitable for real-time detection tasks on resource-constrained devices. While YOLOv5s inherently possesses a well-structured and efficient backbone for feature extraction, the integration of EfficientNetV2’s optimized convolutional layers and scaling mechanisms further enhances the ability to extract discriminative features at multiple spatial scales. This is particularly valuable in complex scenes, where robust feature extraction is essential for detecting and classifying objects with varying appearances, sizes, and backgrounds [17] . To improve detection across object scales, YOLOv5s utilizes a multi-scale feature fusion mechanism—leveraging the strengths of FPN (Feature Pyramid Network) and PANet structures to combine feature maps from different resolutions. The incorporation of EfficientNetV2 complements this design by enhancing low-level and high-level feature representations, thereby improving both localization accuracy and semantic understanding. The synergistic integration of these two networks allows for more precise detection of small, medium, and large targets, and better utilization of hierarchical feature information across multiple receptive fields. The complete architecture of the proposed YOLOv5s_EfficientNetV2 model, including the detailed configuration of each module, is illustrated in Figure 6. 3.3 Visualization Techniques for Bronze Disease Identification 3.3.1 Opencv OpenCV provides robust support for reading image files in various formats, such as JPEG and PNG, and converting them into data structures suitable for further processing. Processed images can be readily saved in user-defined formats, facilitating documentation and sharing of results. Image scaling is often necessary to meet the input requirements of deep learning models or to optimize computational efficiency. To preserve visual quality during resizing, OpenCV supports multiple interpolation techniques. In image recognition tasks, different color spaces offer distinct advantages. For instance, grayscale images are often more efficient for specific feature extraction operations. OpenCV supports a wide range of color space conversions, including transformations from BGR to grayscale, HSV, and others. Additionally, OpenCV offers a comprehensive suite of filtering operations, such as Gaussian, median, and bilateral filtering, which are commonly used to denoise images, enhance features, or smooth textures. Edge detection is a critical step in many image processing pipelines, as edges often correspond to object boundaries. OpenCV includes several effective edge detection algorithms, such as the Canny detector and Sobel operator, which are widely used for contour extraction and object localization. 3.3.2 Pyqt For building graphical user interfaces (GUIs), PyQt5, a Python binding of the Qt library, provides a powerful toolkit for developing cross-platform applications. It includes a comprehensive set of interface components such as buttons, text fields, windows, and sliders, enabling flexible and interactive application design. PyQt5 also supports advanced multimedia handling, making it possible to render images, respond to user inputs (e.g., clicks and drags), and visualize detection results interactively. These features are particularly useful in applications such as bronze disease detection, where visualizing diagnostic outcomes clearly and intuitively is essential. Moreover, PyQt5's cross-platform compatibility across Windows, Linux, and macOS allows seamless deployment of applications, such as lane line inspection systems, across diverse environments to meet varying user demands. 4. Verification of the validity of the model results and the improved method 4.1 Batch-size selection Batch size, a critical hyperparameter in the training process of deep learning models, significantly influences the model's generalization capability, convergence behavior, and training efficiency. In the context of bronze disease recognition and classification using the enhanced YOLOv5 model, this study evaluates the effect of batch size on model performance [ 35 ] . Following the integration of the EfficientNetV2 backbone—a lightweight and efficient enhancement to the original YOLOv5s architecture—a series of experiments were conducted using batch sizes of 2, 4, 8, and 16. The results indicate that the model achieves its best performance at a batch size of 8, with the following metrics: Precision = 0.94%, Recall = 0.884%, mAP50 = 0.916%, and mAP50: 95 = 0.513%, as shown in Table 1 . Table 1 Performance comparison of models with different batch sizes Model Batch-size Precision(%) Recall(%) mAP50(%) mAP50: 95(%) Yolov5s_ EfficientNetV2 2 0.898 0.864 0.872 0.458 4 0.901 0.869 0.886 0.5 8 0.94 0.884 0.916 0.513 16 0.914 0.876 0.911 0.507 These performance values are summarized in Table 1 , and the trends across different batch sizes are illustrated in Fig. 7 . As the batch size increases, model performance initially improves and then declines. This observation suggests that a moderate batch size effectively balances the accuracy and stability of gradient estimation, thereby enhancing the model’s ability to detect and classify bronze diseases more reliably. 4.2 Model Comparison Table 2 presents the experimental results of various models employed in the bronze disease recognition task based on the enhanced YOLOv5s architecture. Notably, the YOLOv5s_EfficientNetV2 model demonstrates significant improvements, with Precision and [email protected] enhanced by 8.0% and 2.5%, respectively, compared to the baseline YOLOv5s model. These improvements indicate that integrating EfficientNetV2 as the backbone network effectively enhances the model's feature extraction capabilities, particularly in capturing the intricate textures and visual complexity associated with bronze disease manifestations. The variant YOLOv5s_EfficientNetV2_CA, which incorporates the Coordinate Attention (CA) mechanism, shows slightly lower performance in Recall and [email protected] compared to its non-CA counterpart. However, it still outperforms the original YOLOv5s model across all metrics. This suggests that while the CA module can suppress background noise and potentially improve robustness, it may also diminish the sensitivity to fine-grained local features, warranting further evaluation of its trade-offs. Overall, the improved YOLOv5s variants demonstrate substantial advantages in bronze disease recognition, offering both high accuracy and robustness, and thereby providing reliable technical support for the intelligent detection of cultural relic diseases. The source code and network structure for each model are provided in Table S3–S8. Table 2 Comparison of detection effect of different algorithms Model Precision Recall mAP50 mAP0.5:0.95 Yolov5s_ EfficientNetV2 0.940 0.884 0.916 0.520 Yolov5s_ EfficientNetV2_CA 0.936 0.883 0.910 0.513 Yolov5s 0.870 0.867 0.894 0.494 Yolov5s_GCB 0.890 0.834 0.888 0.497 The average accuracy curves of the YOLOv5s, YOLOv5s_GCB, YOLOv5s_EfficientNetV2_CA, and YOLOv5s_EfficientNetV2 models are illustrated in Fig. 8 . All models converge after approximately 200 training iterations. Among them, the baseline YOLOv5s achieves an average accuracy of around 87%, while YOLOv5s_EfficientNetV2 achieves the highest accuracy at approximately 92%. The YOLOv5s_GCB model stabilizes at about 82%, and YOLOv5s_EfficientNetV2_CA reaches 87%. These results further validate the effectiveness of EfficientNetV2 in enhancing detection performance for this domain-specific task. 4.3 EfficientNetV2-YOLOV5 modeling results The EfficientNetV2-YOLOv5s network model was constructed following the algorithmic framework proposed in this study. The training and validation loss curves for the model are presented in Fig. 9 . During the initial training phase, both loss values decrease sharply, indicating rapid convergence. After approximately 200 training epochs, the curves stabilize, with the training loss converging around 0.045 and the validation loss around 0.03, suggesting that the model reaches a relatively low and consistent loss level, indicative of good generalization and convergence performance. The performance of the YOLOv5s_EfficientNetV2 model is further evaluated through multiple indicators, including Precision-Recall (PR) curves, [email protected] , [email protected] :0.95, Precision, and Recall, as illustrated in Fig. 10 . The PR curve represents the relationship between recall (x-axis) and precision (y-axis). The area under this curve reflects the mean Average Precision (mAP), and values closer to 1 indicate better model performance. Among the four defect categories, the mineralization category achieves the highest detection accuracy, with an mAP of 0.988, demonstrating the model's strong capability in recognizing this specific lesion type. The mAP scores for cracking and holes are 0.821 and 0.911, respectively, while the incomplete category achieves 0.946, placing its performance between that of holes and mineralization. The model shows high precision at lower recall levels, but as recall increases, the precision begins to decline due to a rise in false positives, a common trade-off in classification tasks. The Confusion Matrix (CM) is shown in Fig. 11 , providing a detailed assessment of classification accuracy for each defect type. The horizontal axis denotes the ground truth labels, while the vertical axis indicates the predicted labels. The categories mineralization, holes, and incomplete exhibit strong classification accuracy. Although the model shows slightly lower performance for the cracking category, it still achieves a recognition accuracy of 84%. Overall, the EfficientNetV2-enhanced YOLOv5s model exhibits robust classification performance across all defect categories, validating its effectiveness for intelligent detection and classification of bronze disease features in cultural heritage artifacts. 4.4 Visualization results Figure 12 presents the visualization interface designed for the intelligent detection of bronze diseases. The user can select a pre-trained model, such as the YOLOv5s_EfficientNetV2 model or the best-performing model, by clicking the "Import Model" button located in the navigation bar. After that, the user can upload an input image, namely, a bronze ware disease image—via the "Insert Image" button. Upon clicking the "Start Detection" button, the system navigates to the Bronze Disease Detection page, where the uploaded image is processed, and the locations and types of bronze disease features are detected and displayed in real time. The detection results for various defect categories are shown in Figs. 13 through 16 , which demonstrate the model’s capability to identify and localize different types of bronze ware deterioration. Specifically: Fig. 13 illustrates the recognition results for mutilation, Fig. 14 shows the detection of holes, Fig. 15 presents results for whole-body mineralization, and Fig. 16 highlights the model’s performance in detecting cracks. These visualizations further validate the practicality and effectiveness of the enhanced model in real-world applications, providing a user-friendly interface for cultural heritage conservation and disease diagnosis. 4.5 Suggestions for the Conservation of Bronze Ware Taking the bronze artifacts housed in the Jingmen Museum as a case study, this research employed the improved YOLOv5s_EfficientNetV2 model to detect and classify surface degradation features. The enhanced model leverages compound scaling and an optimized convolutional structure to significantly improve the efficiency of multi-scale feature extraction, enabling more accurate and robust recognition of disease patterns. Furthermore, a visual interface developed with OpenCV and PyQt5 allows for intuitive visualization of the location, category, and confidence level of detected diseases. This facilitates more informed decision-making for subsequent conservation and restoration processes. Based on the diagnostic results, the following preservation and restoration strategies are proposed: 1. Incomplete Structures For minor surface damage, missing components can be restored using welding techniques, with strict control of the welding temperature to prevent thermal damage to adjacent areas. In cases of large-scale structural loss, riveting or bonding methods are more suitable. The repaired surfaces should undergo artificial aging—via chemical patination or pigment application—to harmonize the color and texture with the original artifact, thereby maintaining both aesthetic integrity and historical authenticity. 2. Holes Begin by using ultrasonic cleaning equipment to remove corrosion products and foreign matter from the holes. Subsequently, reinforce the fragile structures surrounding the holes using high-molecular materials such as epoxy resin or acrylic resin, which can be introduced via injection or brushing to enhance structural strength. Appropriate filling materials—chosen based on the size and geometry of the holes—are then embedded using welding or mechanical fixation (e.g., bolts), followed by surface filling and cosmetic restoration. 3. Whole-body Mineralization For extensively mineralized regions, corrosion suppression should be carried out using chemical inhibitors. A diluted benzotriazole (BTA) solution (typically 0.5–1%) is employed, with immersion time adjusted according to the severity of mineralization. BTA molecules form a protective complex layer on the metal surface, which blocks contact with oxygen and moisture, thereby inhibiting further corrosion. After treatment, apply a silicone-based sealant (e.g., polydimethylsiloxane) via spraying or brushing to form a transparent, breathable film, which both prevents external contamination and allows internal moisture to dissipate—minimizing the risk of localized corrosion due to sealing. 4. Cracks Narrow cracks can be filled with low-viscosity epoxy resin and injected using a pressure syringe. The adhesive’s good fluidity ensures complete penetration and sealing, effectively preventing fissure propagation. For larger cracks (> 0.5 mm), the fissure edges should be cleaned using micro-mechanical tools to remove rust and debris. Then, bronze strips—matching the original material—are inserted and fixed via welding or bonding. The surface is then ground and aged to ensure visual and material continuity with the original artifact. 5. Conclusion In this study, an enhanced YOLOv5s_EfficientNetV2 model was developed to improve the recognition of surface diseases in bronze artifacts. By integrating the lightweight and high-performance EfficientNetV2 backbone, the model demonstrates significant improvements in precision, recall, and multi-scale feature extraction. The optimal batch size was identified as 8, balancing accuracy and computational efficiency. The addition of attention mechanisms further enhanced the model's robustness in complex detection environments. A user-friendly visual interface was also developed using OpenCV and PyQt5, enabling an intuitive display of disease types, locations, and confidence levels. This tool provides practical support for cultural heritage professionals in the protection and restoration of bronze wares. Although the current dataset includes only four disease categories, the model exhibits strong scalability and adaptability. As more annotated data become available, it is expected to achieve even greater accuracy and generalization. In the long term, this method holds promise not only for broader applications in bronze conservation but also for digital preservation efforts across a wide range of cultural heritage artifacts. Declarations Author Contribution Han Wu and Jing Yang wrote the main manuscript text. All authors prepared figures and tables. All authors reviewed the manuscript. Acknowledgments This research was supported by the Open Project of the Key Scientific Research Base of the National Cultural Heritage Administration for the Protection of Unearthed Wood Lacquerware (2021H10198, 2023H10017). References Wang Z, Xi X, Li L, Zhang Z, Han Y, Wang X, Sun Z, Zhao H, Yuan N, Li H, Yan B. Chen. Tracking the progression of the simulated bronze disease—a laboratory X-ray microtomography study. Molecules. 2023;28(13):4933–49. Saraiva AS, Figueiredo E, Águas H, Silva RJC. Characterisation of archaeological high-tin bronze corrosion structures. Stud Conserv. 2022;67(4):222–36. Li J, Li L, Xie Z, Xiang J, Zhao X, Xiao Q. Ling. A comprehensive assessment method for the health status of bronzes unearthed at archaeological sites. Herit Sci. 2023;11(1):86–101. Jiang L, Xia Q, He T, Dong W. Rational construction of a protective layer for corroded bronze based on composite corrosion inhibitors. Herit Sci. 2025;13(1):187–96. Wu M, Yang L, Chai R. Research on multi-scale fusion method for ancient bronze ware X-ray images in NSST domain. Appl Sci. 2024;14(10):4166–84. Liu Y, Zhang S, Li W, Yue H. Application of X-Ray Technology in Detecting the Disease Structure of Bronze Cultural Relics. The 15th National Nuclear Structure Conference, Guilin, Guangxi, China, 20149(in Chinese). Pidaparti RM, Aghazadeh BS, Whitfield A, Bao AS. Mercier. Classification of corrosion defects in NiAl bronze through image analysis. Corros Sci. 2010;52(11):3661–6. Chen Y, Huang Y, Zhang Z, Wang Z, Liu B, Liu C, Huang C, Dong S, Pu X, Wan F, Qiao X. Qian. Plant image recognition with deep learning: A review. Comput Electron Agric. 2023;212:108072–88. He T. Image quality recognition technology based on deep learning. J Vis Commun Image Represent. 2019;65:102654–61. Liu K, Lin K, Zhu C. Research on Chinese traditional opera costume recognition based on improved YOLOv5. Herit Sci. 2023;11(1):1–14. Xia S. Research on Disease Detection Method for BronzeArtifacts based on Deep Learning. Shaanxi: Master's Thesis of Northwest University, 2022(in Chinese). Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int J Comput Vision. 2020;128:336–59. Xie Y, Hu W, Xie S, He L. Surface defect detection algorithm based on feature-enhanced YOLO. Cogn Comput. 2023;15(2):565–79. Zhao Y, Wang H, Xie X, Xie Y, Yang C. An enhanced yolov5-based algorithm for metal surface defect detection. Appl Sci. 2023;13(20):11473–89. Shi J, Yang J, Zhang Y. Research on steel surface defect detection based on YOLOv5 with attention mechanism. Electronics. 2022;11(22):3735–49. Zhao Z, Bakar EBA, Razak NBA. M. N. Akhtar. Corrosion image classification method based on EfficientNetV2. Heliyon. 2024;10(17):1–22. Feng K, He S, Wu X, Jiang P, Huang S. Enhanced YOLOv5 for micro-defect detection on KDP crystal surfaces: a fusion of EfficientNetV2 and normalized Wasserstein distance. J Real-Time Image Proc. 2025;22(2):72–83. Wang Z, Wu L, Li T, Shi P. A smoke detection model based on improved YOLOv5. Mathematics. 2022;10(7):1190–202. Liu B. Luo. An improved Yolov5 for multi-rotor UAV detection. Electronics. 2022;11(15):2330–42. Liang J, Kong R, Ma R, Zhang J, Bian X. Aluminum surface defect detection algorithm based on improved yolov5. Adv Theory Simulations. 2024;7(2):2300695–705. Zeng L, Duan X, Pan Y, Deng M. Research on the algorithm of helmet-wearing detection based on the optimized yolov4. Visual Comput. 2023;39(5):2165–75. Li Y, Wang Y, Sui D, Guo M. Dense Buddha head object detection and counting YOLOv8 network based on multi-scale attention and data augmentation fusion. Herit Sci. 2025;13(1):23–37. Lei X, Pan H, Huang X. A dilated CNN model for image classification. IEEE access. 2019;7:124087–95. Liu G, Hu Y, Chen Z, Guo J, Ni P. Lightweight object detection algorithm for robots with improved YOLOv5. Eng Appl Artif Intell. 2023;123:106217–30. Xiong C, Hu S, Fang Z. Application of improved YOLOV5 in plate defect detection. Int J Adv Manuf Technol, 2022: 1–13. Lang X, Ren Z, Wan D, Zhang Y, Shu S. MR-YOLO: An improved YOLOv5 network for detecting magnetic ring surface defects. Sensors. 2022;22(24):9897–913. He K, Zhang X, Ren S, Sun J. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans Pattern Anal Mach Intell. 2015;37(9):1904–16. Li F, Xiao K, Hu Z, Zhang G. Fabric defect detection algorithm based on improved YOLOv5. Visual Comput. 2024;40(4):2309–24. Liu S, Qi L, Qin H, Shi J, Jia J. Path aggregation network for instance segmentation. Proceedings of the IEEE conference on computer vision and pattern recognition, 2018: 8759–8768. Zhang J, Zhang Y, Liu J, Lan Y, Zhang T. Human figure detection in Han portrait stone images via enhanced YOLO-v5. Herit Sci. 2024;12(1):1–16. Tan M, Le Q. Efficientnet: Rethinking model scaling for convolutional neural networks. International conference on machine learning. PMLR, 2019: 6105–6114. Sifre L, Mallat S. Rigid-motion scattering for texture classification. arxiv preprint arxiv:1403 1687, 2014: 1–19. Tan M. Q. Le. Efficientnetv2: Smaller models and faster training. International conference on machine learning, PMLR, 2021: 10096–10106. Kanai S, Fujiwara Y, Yamanaka Y. S. Asachi. Sigsoftmax: Reanalysis of the softmax bottleneck. Adv Neural Inf Process Syst, 2018, 31–45. Hwang JS, Lee SS, Gil JW, Lee CK. Determination of optimal batch size of deep learning models with time series data. Sustainability. 2024;16(14):5936–46. Additional Declarations No competing interests reported. Supplementary Files Supportinginformation6.15.docx Supplementary Material Table S1 summarizes the label information corresponding to the training set images. Table S2 the label information corresponding to the validation set images. Table S3 collates the YOLOv5 code edited from PyCharm2021.3.1. Table S4 collates yolov5s-GCB codes. Table S5 organizes the codes for the yolov5s_EfficientNetV2_CA. Table S6 organizes the yolov5s_EfficientNetV2 codes. Table S7 is the common.py code that adds the network structure through PyCharm2021.3.1. Table S8 is use PyCharm2021.3.1 to add the yolo.py code of the network structure. Document 1 summarizes photographs of bronzes and cropped disease maps. 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-6900599","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":473425509,"identity":"44f4624d-34f1-4f59-b5b4-adf6f850c2e7","order_by":0,"name":"Han Wu","email":"","orcid":"","institution":"Jingmen Museum","correspondingAuthor":false,"prefix":"","firstName":"Han","middleName":"","lastName":"Wu","suffix":""},{"id":473425510,"identity":"4e85d79c-d543-4971-ada5-f276fc576b42","order_by":1,"name":"Jing Yang","email":"","orcid":"","institution":"Wuhan University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Yang","suffix":""},{"id":473425511,"identity":"b2ac0762-fe77-4066-9af3-52977a2f02bb","order_by":2,"name":"Lijun Pang","email":"","orcid":"","institution":"Wuhan University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Lijun","middleName":"","lastName":"Pang","suffix":""},{"id":473425512,"identity":"f6980812-dd2e-4e0a-b246-d6ebef6fde13","order_by":3,"name":"Wei Zhou","email":"","orcid":"","institution":"Jingmen Museum","correspondingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Zhou","suffix":""},{"id":473425513,"identity":"a5290a61-89b9-48fb-88fb-7f4b0ddac81d","order_by":4,"name":"Yang Lei","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5ElEQVRIiWNgGAWjYBACPgbGBhAtB+UzE9bCBtViDMQgFlFaICCxgXgt7M1tEj931KZvuJH+/AFDhXViA/vZA/i18Bxsk+w9czx3w40cwwaGM+mJDTx5Cfi1SCS2SfC2HcvddiOHsYGx7XBigwSPAX4t8g/bJP+2HUs3u5H+sIHxHzFaJBjbpHnbahLMbiQYNjA2EKOFJ7HZWrbtgOH+M28MZyQcSzdu48nBr4Wf/fjDm2/b6uQl29MffPhQYy3bz34GvxYgYJFgYDgMYSYwIGIKH2D+wMBQR4S6UTAKRsEoGLEAABwGRtDT9F4xAAAAAElFTkSuQmCC","orcid":"","institution":"Wuhan University of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Yang","middleName":"","lastName":"Lei","suffix":""},{"id":473425514,"identity":"61d9b928-27e1-4d0a-a29a-7987e55ecd9b","order_by":5,"name":"Xuegang Liu","email":"","orcid":"","institution":"Jingzhou Conservation Center","correspondingAuthor":false,"prefix":"","firstName":"Xuegang","middleName":"","lastName":"Liu","suffix":""},{"id":473425515,"identity":"ff4d9af7-0efc-4524-a6e1-606fe707608b","order_by":6,"name":"Shan Ding","email":"","orcid":"","institution":"Jingzhou Conservation Center","correspondingAuthor":false,"prefix":"","firstName":"Shan","middleName":"","lastName":"Ding","suffix":""},{"id":473425516,"identity":"4819518c-7b22-473c-a7b6-992b2386df7f","order_by":7,"name":"Yuqiu Chen","email":"","orcid":"","institution":"University of Delaware","correspondingAuthor":false,"prefix":"","firstName":"Yuqiu","middleName":"","lastName":"Chen","suffix":""}],"badges":[],"createdAt":"2025-06-16 01:23:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6900599/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6900599/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":85176246,"identity":"07a08bc4-547a-4e52-b049-57edee6b7f00","added_by":"auto","created_at":"2025-06-23 06:35:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":46625,"visible":true,"origin":"","legend":"\u003cp\u003eModeling strategy of the disease identification model\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/30c351320eede33c976c0eb9.png"},{"id":85175696,"identity":"02b8b06b-4268-49d5-ae73-9f4225dc0a45","added_by":"auto","created_at":"2025-06-23 06:27:48","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":211999,"visible":true,"origin":"","legend":"\u003cp\u003eExample diagram of typical diseases (a) Whole-body mineralization, (b) cracks, (c) holes, (d) incomplete\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/f34145589a12c8fde6d884b6.png"},{"id":85175710,"identity":"d3910e29-6497-4c7f-b0da-84fb991cba5f","added_by":"auto","created_at":"2025-06-23 06:27:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":15269,"visible":true,"origin":"","legend":"\u003cp\u003eConvolutional neural network pooling layer structure\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/d5acde8cc7c6d1b257514e8c.png"},{"id":85175730,"identity":"57dc6320-3f2e-444e-8dfd-3a78b0e064ee","added_by":"auto","created_at":"2025-06-23 06:27:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":49054,"visible":true,"origin":"","legend":"\u003cp\u003eYOLOv5s flowchart\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/5f9bc6764a91c9bdd683e3cf.png"},{"id":85175700,"identity":"23d91135-df60-4f3c-bf56-f35d887c4693","added_by":"auto","created_at":"2025-06-23 06:27:48","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":20043,"visible":true,"origin":"","legend":"\u003cp\u003eStructure of MBConv and Fused-MBConv\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/08249aac3e54b643d51a65eb.png"},{"id":85176261,"identity":"71a387f7-3f39-457c-b142-da670cdd45b7","added_by":"auto","created_at":"2025-06-23 06:35:49","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":40179,"visible":true,"origin":"","legend":"\u003cp\u003eNetwork diagram of the improved YOLOv5 structure\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/693e6c09a69b7ce6d5f11bf5.png"},{"id":85175708,"identity":"b1976888-4253-44fd-8956-90771966a72c","added_by":"auto","created_at":"2025-06-23 06:27:48","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":82242,"visible":true,"origin":"","legend":"\u003cp\u003ePrecision (a), Recall (b), mAP50 (c), mAP0.5: 0.95 (d) for different Batch-sizes of the model\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/25915bc0e1d1e1df668ef8bb.png"},{"id":85175701,"identity":"5e192765-23b1-495e-995f-9edabf4c0122","added_by":"auto","created_at":"2025-06-23 06:27:48","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":24090,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of accuracy values of various networks\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/467bf37d2ecdc4f4b76e0042.png"},{"id":85176242,"identity":"20750d79-c69d-4e89-89ca-e310015de0e3","added_by":"auto","created_at":"2025-06-23 06:35:48","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":37691,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and validation loss curves\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/f2c5550a1b94146957354e1b.png"},{"id":85175719,"identity":"b753caa2-3161-419a-ab50-7711c6bb8c00","added_by":"auto","created_at":"2025-06-23 06:27:49","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":62433,"visible":true,"origin":"","legend":"\u003cp\u003ePR curves for Yolov5s_ EfficientNetV2\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/aa2b7860a346fc212fccb0f5.png"},{"id":85175718,"identity":"7dab149a-cd36-4cce-9a78-1548ea25ca26","added_by":"auto","created_at":"2025-06-23 06:27:49","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":95251,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix results for improved YOLOv5 algorithm\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/50953ae7156075e1a2144a4a.png"},{"id":85175723,"identity":"c6487b16-38af-4862-8a9b-2e0e9fd2dc07","added_by":"auto","created_at":"2025-06-23 06:27:49","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":46716,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization Interface\u003c/p\u003e","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/e68f50bf8330c161ed607eda.png"},{"id":85175747,"identity":"9ee91eee-1492-43d1-b6f8-a34f2e65792b","added_by":"auto","created_at":"2025-06-23 06:27:50","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":198757,"visible":true,"origin":"","legend":"\u003cp\u003eResults of mutilation detection\u003c/p\u003e","description":"","filename":"13.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/3aaad14ad1cfe0db21d012da.png"},{"id":85175697,"identity":"15fee1f2-4c54-4b5c-a2f7-f21ead0048fd","added_by":"auto","created_at":"2025-06-23 06:27:48","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":234540,"visible":true,"origin":"","legend":"\u003cp\u003eHole detection results\u003c/p\u003e","description":"","filename":"14.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/fbd48f0d4c6712ba33964c27.png"},{"id":85175727,"identity":"b85b2aa5-1073-4c74-b845-0326ac80f94b","added_by":"auto","created_at":"2025-06-23 06:27:49","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":232162,"visible":true,"origin":"","legend":"\u003cp\u003eWhole-body mineralization test results\u003c/p\u003e","description":"","filename":"15.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/a4d842c06bc8595cb3033d90.png"},{"id":85176290,"identity":"77de4038-9769-4608-afc9-e82e17b3b3f6","added_by":"auto","created_at":"2025-06-23 06:35:50","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":259250,"visible":true,"origin":"","legend":"\u003cp\u003eCracks detection results\u003c/p\u003e","description":"","filename":"16.png","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/401edba3c5b46a3aa27a67db.png"},{"id":104397824,"identity":"c9f41a03-527c-4ec6-a2b0-7a0588fe146d","added_by":"auto","created_at":"2026-03-11 11:57:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2406656,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/7bcf0dc5-b5b9-4c0a-8cd6-cf54b1080dde.pdf"},{"id":85175740,"identity":"795e90f4-6976-4adb-9988-75938439f3a8","added_by":"auto","created_at":"2025-06-23 06:27:50","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":34283959,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Material Table S1 summarizes the label information corresponding to the training set images. Table S2 the label information corresponding to the validation set images. Table S3 collates the YOLOv5 code edited from PyCharm2021.3.1. Table S4 collates yolov5s-GCB codes. Table S5 organizes the codes for the yolov5s_EfficientNetV2_CA. Table S6 organizes the yolov5s_EfficientNetV2 codes. Table S7 is the common.py code that adds the network structure through PyCharm2021.3.1. Table S8 is use PyCharm2021.3.1 to add the yolo.py code of the network structure. Document 1 summarizes photographs of bronzes and cropped disease maps.\u003c/p\u003e","description":"","filename":"Supportinginformation6.15.docx","url":"https://assets-eu.researchsquare.com/files/rs-6900599/v1/03675196359b85c570c0a2cf.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Identification and Classification of Bronze Surface Diseases Based on Improved YOLOv5","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBronzes hold profound historical, cultural, and artistic value and represent a significant component of China\u0026rsquo;s cultural heritage. However, over time and under the influence of environmental factors, various forms of corrosion, both on the surface and within the structure, have developed\u003csup\u003e[\u003cspan additionalcitationids=\"CR2 CR3\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/sup\u003e. Common types of deterioration include mutilation, porosity, full-body mineralization, and fractures. These not only compromise the aesthetic appeal of the artifacts but also pose serious risks to their structural integrity\u003csup\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e. Without timely identification and intervention, such diseases can spread, leading to irreversible damage. Traditional manual inspection methods heavily rely on experience and are prone to subjective error, often resulting in suboptimal restoration strategies. Deep learning-based object detection technologies offer promising solutions by automatically identifying disease boundaries, classifying and locating defect regions, and outputting structured data including coordinates and confidence scores. This capability makes such technologies highly valuable for accurate and timely diagnosis of bronze disease.\u003c/p\u003e \u003cp\u003eConventional detection methods largely depend on visual assessment and expert judgment. These approaches are inefficient, subjective, and inadequate for large-scale artifact conservation. In response, modern detection technologies have been gradually introduced. For example, Liu et al.\u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e employed dual-energy X-ray imaging to reconstruct 3D structures and detect internal damage in bronzes. Ramana M. et al.\u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e utilized wavelet transforms and Fourier analysis on SEM images to quantify pitting and cracking in Ni\u0026ndash;Al bronze surfaces by estimating fractal dimensions. While effective, these methods require substantial computational resources and are often based on linear classification, limiting their precision in complex or overlapping cases. Recent advancements in computer vision and deep learning have significantly enhanced performance in image recognition and classification, particularly with complex, high-dimensional data\u003csup\u003e[\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/sup\u003e. These techniques not only boost detection efficiency but also reduce human error, making them well-suited for the preventive preservation of bronzes. Some research efforts have already applied deep learning to bronze disease identification. For instance, Xia \u003csup\u003e[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e proposed a method using an improved DenseNet architecture combined with Class Activation Mapping (CAM) to achieve weakly supervised disease localization. By integrating SE Blocks and the CBAM attention module, the model effectively extracted and identified disease features. However, this approach still has limitations: weakly supervised localization methods like ES-CAM, while superior to traditional techniques, rely heavily on feature pyramids and complex post-processing steps with high computational costs, making real-time application challenging. Furthermore, the need for extensive architectural modification complicates implementation\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e. In contrast, YOLOv5\u0026mdash;a lightweight object detection model known for its fast inference and high detection accuracy\u0026mdash;is better suited for real-world disease detection scenarios. Xie et al.\u003csup\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e introduced a feature-enhanced YOLO framework for detecting surface defects, improving FPN design and spatial correlation at multiple scales. Similarly, Zhao et al.\u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e proposed an enhanced YOLOv5 model for metal surface defect detection, integrating a parameter-free spatial attention (PSA) mechanism that improved feature extraction from key regions and boosted defect detection performance by 2.4% over the baseline. Despite these improvements, the original YOLOv5 model still struggles with dense object detection and small-scale features\u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn response to these challenges, this study presents an enhanced YOLOv5s model for detecting and classifying surface diseases in bronze wares, aiming to address the limitations of traditional methods in terms of efficiency and objectivity. The model integrates the lightweight and high-performance EfficientNetV2 backbone, which enhances detection of small and dense targets. By incorporating a compound scaling method and optimized convolution structure, the model improves multi-scale feature extraction while reducing computational cost and parameter size\u003csup\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e. To evaluate the model's effectiveness, high-resolution images of bronze artifacts from the Jingmen Museum collection were used to create a dataset. A total of 1,037 images (600\u0026times;600 pixels) were labeled with disease categories such as flaws, holes, full-body mineralization, and cracks. After data augmentation, the dataset was split into training and test sets in an 8:2 ratio. The performance of the improved model was compared against the baseline YOLOv5s and an attention-enhanced variant by analyzing key metrics\u0026mdash;Precision, Recall, and [email protected]. Additionally, the effect of batch size on model convergence and generalization was investigated. To support practical application, a visual detection interface was developed using OpenCV and PyQt5, enabling automatic disease classification and intuitive spatial visualization of detection results. This provides intelligent technical support for the preservation and restoration of bronze wares. Moreover, the proposed methodology can be extended to other domains involving cultural heritage diagnostics, underscoring its broad practical relevance.\u003c/p\u003e"},{"header":"2. Modeling methods and steps","content":"\u003cp\u003eTo address the limitations of traditional manual detection methods for bronze ware diseases, characterized by low efficiency and high error rates, this study focuses on the bronze artifacts housed in the Jingmen Museum and constructs a dedicated dataset for automated analysis. High-resolution surface images of bronze wares were captured using a Canon G1X Mark III digital camera. In accordance with national classification standards for cultural heritage deterioration, images depicting four typical types of disease, whole-body mineralization, cracks, holes, and surface defects, were selected for further analysis. Preprocessing steps, including uniform cropping to a resolution of 600\u0026times;600 pixels, rotation, and flipping, were carried out using Adobe Photoshop. To alleviate the problem of overfitting caused by a limited sample size, data augmentation techniques such as random cropping and translation were applied, expanding the dataset to 1,037 images.\u003c/p\u003e \u003cp\u003eDisease regions were annotated using the LabelImg tool, generating YOLO-format annotation files containing both disease categories and bounding box coordinates. To improve the original YOLOv5s model\u0026rsquo;s limitations in small object detection and multi-scale feature fusion, this study integrates EfficientNetV2 as a lightweight backbone network. Model development and training were conducted in PyCharm using the PyTorch framework, with a batch size of 8 determined empirically. To provide an intuitive visualization of the detection results, an interactive GUI application was developed by combining OpenCV and PyQt5. OpenCV handles image preprocessing, bounding box drawing, and category labeling, while PyQt5 is used to construct the interface, offering functionalities such as model import, image upload, and real-time detection. The system enables the localization and classification of diseases in input images, with results presented via bounding boxes and confidence scores. In the model validation phase, detection performance was quantitatively evaluated using standard metrics, including Precision, Recall, and mean Average Precision at IoU thresholds of 0.5 (mAP50) and from 0.5 to 0.95 (mAP50:95). The overall model architecture and workflow are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data sets and pre-processing\u003c/h2\u003e \u003cp\u003eThis study uses bronze wares from the Jingmen Museum collection as representative specimens. High-resolution digital photographs of the artifacts were captured using a Canon G1X Mark III camera. To accommodate the complex and varied surface lighting conditions of the bronze wares, adjustable parameters such as aperture, shutter speed, and ISO were carefully configured. Specifically, an aperture (F-number) between F5.6 and F8 was chosen to strike a balance between image sharpness and depth of field. An excessively small aperture can cause diffraction-related image degradation, while a large aperture may result in loss of detail due to a shallow depth of field. The shutter speed was controlled within the range of 1/100 to 1/200 seconds to reduce motion blur caused by handheld shaking. In low-light conditions, the shutter speed could be slightly reduced, provided the camera was stabilized (e.g., with a tripod), and the lowest possible ISO setting was used to minimize image noise. Image preprocessing, including random cropping, flipping, and rotation, was conducted using Adobe Photoshop. All images were exported in JPG format with a standardized resolution of 600\u0026times;600 pixels, resulting in a total of 1,037 annotated images depicting four distinct types of bronzeware surface defects. The finalized cropped dataset is available in the Supporting Information (Document S1). LabelImg was used to annotate each image by marking the diseased regions and assigning defect categories. Annotations were saved in the YOLO format, containing information such as the bounding box\u0026rsquo;s center coordinates (x, y) and its width (W) and height (H). A summary of the annotation results for each image is presented in Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u0026ndash;S2. The dataset was randomly divided into training and prediction (validation) sets in a ratio of 8:2. Representative samples of annotated disease regions are illustrated in Fig.\u0026nbsp;2.\u003c/p\u003e\u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Relevant evaluation indicators\u003c/h2\u003e \u003cp\u003eIn this study, the performance of the model is evaluated using four key metrics: Precision, Recall, Average Precision (AP), and mean Average Precision (mAP) at different Intersection over Union (IoU) thresholds, specifically from 0.50 to 0.95. Equations\u0026nbsp;(1) and (2) define Precision and Recall, respectively\u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/sup\u003e.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:Precision=\\frac{TP}{TP+FP}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(1\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:Recall=\\frac{TP}{TP+FN}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(2\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere TP refers to the number of instances correctly identified as positive (i.e., correctly detected diseased regions). FP denotes the number of instances incorrectly predicted as positive (i.e., non-diseased areas mistakenly labeled as diseased). FN represents the number of actual positive instances that the model fails to detect (i.e., diseased areas predicted as healthy)\u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/sup\u003e. Higher Precision indicates that fewer false positives are present in the predictions, while higher Recall reflects the model\u0026rsquo;s ability to detect the majority of actual positive samples.\u003c/p\u003e \u003cp\u003eEquations\u0026nbsp;(3) and (4) define AP and mean Average Precision ([email protected])\u003csup\u003e[\u003cspan additionalcitationids=\"CR21\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e.\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:AP={\\int\\:}_{0}^{1}P\\left(R\\right)dR\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(3\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:mAP=\\frac{{\\sum\\:}_{i=1}^{n}A{P}_{i}}{n}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(4\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe AP for a single class is computed as the area under the Precision\u0026ndash;Recall (P\u0026ndash;R) curve, and mAP is the mean of AP values across all classes, where \u0026#119899; is the number of categories considered. In this experiment, [email protected] and [email protected]:0.95 are both reported to provide a comprehensive assessment of the model\u0026rsquo;s detection performance across varying IoU thresholds.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. YOLOv5 structure and improvements","content":"\u003ch2\u003e3.1 Convolutional Network Neural\u003c/h2\u003e\n\u003cp\u003eConvolutional Neural Networks (CNNs) are widely used deep learning algorithms particularly well-suited for processing multidimensional data such as images, speech, and natural language. Their key attributes, local connectivity and parameter sharing, enable them to recognize image features both quickly and accurately, making CNNs highly effective across various tasks and application domains. A typical CNN architecture is composed of five fundamental components: the input layer, convolutional layers, pooling layers, fully connected layers, and the output layer \u003csup\u003e[23]\u003c/sup\u003e. CNNs operate by alternating between convolutional, pooling, and fully connected layers, which allows them to automatically extract high-level, abstract features from input data. The globally shared weights and locally connected structure of CNNs reduce the number of learnable parameters and enhance computational efficiency, making them particularly adept at handling high-dimensional inputs such as images.\u003c/p\u003e\n\u003cp\u003eAmong these components, the convolutional layer plays a central role in feature extraction. It processes the input image using a set of learnable filters (convolutional kernels), which perform localized weighted summations followed by the addition of a bias term. Each kernel is designed to detect specific local features such as edges, textures, or color patterns, enabling hierarchical feature learning as the network depth increases. Following the convolutional layer is the pooling layer, which primarily serves to reduce the spatial dimensions of the data while retaining essential feature information. As illustrated in Figure 3, the most commonly used pooling operations are max pooling and average pooling. Max pooling selects the highest value within a defined window, while average pooling computes the mean of values in that region. Pooling not only reduces the computational burden but also enhances the model\u0026rsquo;s robustness to small shifts, noise, and distortions in the input. Finally, the fully connected (FC) layer appears near the end of the network and is responsible for mapping the extracted features to the target output space, such as classification labels. Each neuron in the FC layer is connected to every neuron in the preceding layer, allowing the model to integrate spatially distributed features and generate the final prediction.\u003c/p\u003e\n\u003ch2\u003e3.2 Improved YOLOv5s network architecture\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003e3.2.1 YOLOv5s\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eYOLOv5s, a lightweight variant within the YOLO (You Only Look Once) series of object detection models, is widely favored for its efficiency, accuracy, and ease of deployment. Its network architecture is primarily composed of three components: the Backbone, Neck, and Head layers\u003csup\u003e[24, 25]\u003c/sup\u003e. These components work in conjunction to perform feature extraction, feature fusion, and target detection from the input image. The overall structure of the model is illustrated in Figure 4.\u003c/p\u003e\n\u003cp\u003eThe Backbone is responsible for receiving and preprocessing the input image. Within the Backbone, the Focus module performs a slicing operation on the input image, followed by convolutional processing to extract features. By reducing the number of channels and increasing the spatial resolution (i.e., width and height) of the feature map, the Focus module preserves more detailed spatial information.\u003csup\u003e[26]\u003c/sup\u003e. Another critical component of the Backbone is the CSPNet (Cross Stage Partial Network) module, a lightweight structure designed to improve the learning ability of the network while reducing computational cost\u003csup\u003e[20]\u003c/sup\u003e. Additionally, the SPP (Spatial Pyramid Pooling) module enhances the detection of objects at multiple scales by performing pooling operations at different receptive field sizes\u003csup\u003e[27, 28]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThe Neck of YOLOv5s incorporates both FPN (Feature Pyramid Network) and PANet (Path Aggregation Network) structures. FPN generates multi-scale feature maps by combining features across different levels via lateral connections and up-sampling operations \u003csup\u003e[28]\u003c/sup\u003e. PANet further strengthens this information flow by introducing bottom-up paths with down-sampling operations, thus improving the fusion of low-level and high-level feature representations\u003csup\u003e[29, 30]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eFollowing feature fusion in the Neck, the Head layer processes the resulting feature maps via convolutional operations to produce final detection outputs. These outputs include bounding box coordinates, objectness scores (confidence levels), and class probabilities, enabling precise target localization and classification. To enhance the model\u0026apos;s generalization performance during training, YOLOv5s employs a series of data augmentation techniques such as mosaic augmentation and adaptive anchor box computation. Furthermore, the model\u0026apos;s lightweight design significantly reduces its computational cost and number of parameters, making it highly suitable for real-time applications on edge devices.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2.2 EfficientNetV2\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn 2019, Tan and Le et al.\u003csup\u003e[31]\u003c/sup\u003e proposed EfficientNet, a family of convolutional neural networks that introduced a novel model scaling approach to achieve a better trade-off between accuracy and efficiency. Their study systematically investigated how scaling the depth, width, and resolution of a network influences its performance. Unlike traditional scaling strategies that focus on a single dimension, EfficientNet adopts a composite scaling method that uniformly scales all three dimensions using a fixed set of scaling coefficients. This technique mitigates the limitations associated with one-dimensional scaling approaches, enabling EfficientNet to achieve state-of-the-art results with significantly fewer parameters and FLOPs. However, EfficientNetV1 exhibited certain limitations during training. For instance, training on large-resolution images was found to be slow and inefficient, particularly in the early layers of the network that employed depthwise separable convolutions\u003csup\u003e[32]\u003c/sup\u003e. While depthwise convolutions reduce parameter count and computational load, they are often not optimized on current hardware accelerators, resulting in suboptimal performance. Moreover, EfficientNetV1\u0026apos;s approach of uniformly scaling all network stages was later shown to be inefficient, as different stages contribute unequally to overall model complexity and training dynamics. To address these issues, EfficientNetV2 introduced several architectural improvements, notably the Fused-MBConv block\u003csup\u003e[33]\u003c/sup\u003e. In contrast to the original MBConv structure\u0026mdash;comprising an expansion 1\u0026times;1 convolution followed by a depthwise 3\u0026times;3 convolution\u0026mdash;the Fused-MBConv simplifies this pipeline by replacing it with a standard 3\u0026times;3 convolution, enhancing compatibility with modern mobile and server accelerators (as illustrated in Figure 5).\u003c/p\u003e\n\u003cp\u003eAnother key innovation in EfficientNetV2 is its non-uniform scaling strategy, which scales different network stages unequally based on their contribution to model efficiency and training time. This adjustment improves both training speed and parameter utilization. The network begins with standard preprocessing operations, such as normalization and resizing, to prepare the input images. Feature extraction is then performed through successive convolutional and pooling layers, with many of the convolutional layers utilizing depthwise separable convolutions to further reduce computational overhead.\u003c/p\u003e\n\u003cp\u003eTo ensure consistent performance across input images of varying resolutions, a feature deflation module is introduced. This module applies global average pooling to convert spatial feature maps into fixed-length vectors, which are then passed through a learnable 1\u0026times;1 convolution (or linear transformation) to align the feature dimensions. To enhance the network\u0026rsquo;s representational capacity, multi-scale feature fusion strategies, such as feature cascading or aggregation, are applied to combine feature maps from different depths\u003csup\u003e[34]\u003c/sup\u003e. The final stage of the network is typically a fully connected layer (or a final convolutional layer, depending on the task), which outputs either a probability distribution over class labels (using the softmax activation function for classification tasks) or continuous values such as bounding box coordinates for regression tasks. The training process involves computing a loss function-typically cross-entropy for classification or mean squared error (MSE) for regression-based on the discrepancy between predicted outputs and ground truth labels. Network parameters are then iteratively updated using backpropagation to minimize this loss, thereby optimizing the model\u0026rsquo;s predictive performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2.3 YOLOv5s_EfficientNetV2\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study proposes the integration of EfficientNetV2 into the YOLOv5s framework, forming a YOLOv5s_EfficientNetV2 hybrid network to enhance the baseline model\u0026rsquo;s feature extraction capability. EfficientNetV2, with its lightweight architecture and composite scaling strategy, is designed to balance model size, computational efficiency, and representational power. By optimizing the network structure and parameters, EfficientNetV2 is able to retain strong feature extraction performance while reducing the computational burden\u0026mdash;making it highly suitable for real-time detection tasks on resource-constrained devices. While YOLOv5s inherently possesses a well-structured and efficient backbone for feature extraction, the integration of EfficientNetV2\u0026rsquo;s optimized convolutional layers and scaling mechanisms further enhances the ability to extract discriminative features at multiple spatial scales. This is particularly valuable in complex scenes, where robust feature extraction is essential for detecting and classifying objects with varying appearances, sizes, and backgrounds \u003csup\u003e[17]\u003c/sup\u003e. To improve detection across object scales, YOLOv5s utilizes a multi-scale feature fusion mechanism\u0026mdash;leveraging the strengths of FPN (Feature Pyramid Network) and PANet structures to combine feature maps from different resolutions. The incorporation of EfficientNetV2 complements this design by enhancing low-level and high-level feature representations, thereby improving both localization accuracy and semantic understanding. The synergistic integration of these two networks allows for more precise detection of small, medium, and large targets, and better utilization of hierarchical feature information across multiple receptive fields.\u003c/p\u003e\n\u003cp\u003eThe complete architecture of the proposed YOLOv5s_EfficientNetV2 model, including the detailed configuration of each module, is illustrated in Figure 6.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e3.3 Visualization Techniques for Bronze Disease Identification\u003c/h2\u003e\n\u003cp\u003e\u003cstrong\u003e3.3.1 Opencv\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOpenCV provides robust support for reading image files in various formats, such as JPEG and PNG, and converting them into data structures suitable for further processing. Processed images can be readily saved in user-defined formats, facilitating documentation and sharing of results. Image scaling is often necessary to meet the input requirements of deep learning models or to optimize computational efficiency. To preserve visual quality during resizing, OpenCV supports multiple interpolation techniques. In image recognition tasks, different color spaces offer distinct advantages. For instance, grayscale images are often more efficient for specific feature extraction operations. OpenCV supports a wide range of color space conversions, including transformations from BGR to grayscale, HSV, and others. Additionally, OpenCV offers a comprehensive suite of filtering operations, such as Gaussian, median, and bilateral filtering, which are commonly used to denoise images, enhance features, or smooth textures. Edge detection is a critical step in many image processing pipelines, as edges often correspond to object boundaries. OpenCV includes several effective edge detection algorithms, such as the Canny detector and Sobel operator, which are widely used for contour extraction and object localization.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3.2 Pyqt\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor building graphical user interfaces (GUIs), PyQt5, a Python binding of the Qt library, provides a powerful toolkit for developing cross-platform applications. It includes a comprehensive set of interface components such as buttons, text fields, windows, and sliders, enabling flexible and interactive application design. PyQt5 also supports advanced multimedia handling, making it possible to render images, respond to user inputs (e.g., clicks and drags), and visualize detection results interactively. These features are particularly useful in applications such as bronze disease detection, where visualizing diagnostic outcomes clearly and intuitively is essential. Moreover, PyQt5\u0026apos;s cross-platform compatibility across Windows, Linux, and macOS allows seamless deployment of applications, such as lane line inspection systems, across diverse environments to meet varying user demands.\u003c/p\u003e"},{"header":"4. Verification of the validity of the model results and the improved method","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Batch-size selection\u003c/h2\u003e \u003cp\u003eBatch size, a critical hyperparameter in the training process of deep learning models, significantly influences the model's generalization capability, convergence behavior, and training efficiency. In the context of bronze disease recognition and classification using the enhanced YOLOv5 model, this study evaluates the effect of batch size on model performance\u003csup\u003e[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]\u003c/sup\u003e. Following the integration of the EfficientNetV2 backbone\u0026mdash;a lightweight and efficient enhancement to the original YOLOv5s architecture\u0026mdash;a series of experiments were conducted using batch sizes of 2, 4, 8, and 16. The results indicate that the model achieves its best performance at a batch size of 8, with the following metrics: Precision\u0026thinsp;=\u0026thinsp;0.94%, Recall\u0026thinsp;=\u0026thinsp;0.884%, mAP50\u0026thinsp;=\u0026thinsp;0.916%, and mAP50: 95\u0026thinsp;=\u0026thinsp;0.513%, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance comparison of models with different batch sizes\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBatch-size\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision(%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall(%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003emAP50(%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003emAP50: 95(%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eYolov5s_ EfficientNetV2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.898\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.872\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.458\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.901\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.869\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.886\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.884\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.916\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.513\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.914\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.876\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.911\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.507\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\u003eThese performance values are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, and the trends across different batch sizes are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e7\u003c/span\u003e. As the batch size increases, model performance initially improves and then declines. This observation suggests that a moderate batch size effectively balances the accuracy and stability of gradient estimation, thereby enhancing the model\u0026rsquo;s ability to detect and classify bronze diseases more reliably.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Model Comparison\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the experimental results of various models employed in the bronze disease recognition task based on the enhanced YOLOv5s architecture. Notably, the YOLOv5s_EfficientNetV2 model demonstrates significant improvements, with Precision and [email protected] enhanced by 8.0% and 2.5%, respectively, compared to the baseline YOLOv5s model. These improvements indicate that integrating EfficientNetV2 as the backbone network effectively enhances the model's feature extraction capabilities, particularly in capturing the intricate textures and visual complexity associated with bronze disease manifestations. The variant YOLOv5s_EfficientNetV2_CA, which incorporates the Coordinate Attention (CA) mechanism, shows slightly lower performance in Recall and [email protected] compared to its non-CA counterpart. However, it still outperforms the original YOLOv5s model across all metrics. This suggests that while the CA module can suppress background noise and potentially improve robustness, it may also diminish the sensitivity to fine-grained local features, warranting further evaluation of its trade-offs. Overall, the improved YOLOv5s variants demonstrate substantial advantages in bronze disease recognition, offering both high accuracy and robustness, and thereby providing reliable technical support for the intelligent detection of cultural relic diseases. The source code and network structure for each model are provided in Table S3\u0026ndash;S8.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of detection effect of different algorithms\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003emAP50\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003emAP0.5:0.95\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYolov5s_ EfficientNetV2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.940\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.884\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.916\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.520\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYolov5s_ EfficientNetV2_CA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.936\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.883\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.910\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.513\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYolov5s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.870\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.867\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.894\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.494\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYolov5s_GCB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.890\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.834\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.888\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.497\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 average accuracy curves of the YOLOv5s, YOLOv5s_GCB, YOLOv5s_EfficientNetV2_CA, and YOLOv5s_EfficientNetV2 models are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e8\u003c/span\u003e. All models converge after approximately 200 training iterations. Among them, the baseline YOLOv5s achieves an average accuracy of around 87%, while YOLOv5s_EfficientNetV2 achieves the highest accuracy at approximately 92%. The YOLOv5s_GCB model stabilizes at about 82%, and YOLOv5s_EfficientNetV2_CA reaches 87%. These results further validate the effectiveness of EfficientNetV2 in enhancing detection performance for this domain-specific task.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.3 EfficientNetV2-YOLOV5 modeling results\u003c/h2\u003e \u003cp\u003eThe EfficientNetV2-YOLOv5s network model was constructed following the algorithmic framework proposed in this study. The training and validation loss curves for the model are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e9\u003c/span\u003e. During the initial training phase, both loss values decrease sharply, indicating rapid convergence. After approximately 200 training epochs, the curves stabilize, with the training loss converging around 0.045 and the validation loss around 0.03, suggesting that the model reaches a relatively low and consistent loss level, indicative of good generalization and convergence performance.\u003c/p\u003e\u003cp\u003eThe performance of the YOLOv5s_EfficientNetV2 model is further evaluated through multiple indicators, including Precision-Recall (PR) curves, [email protected], [email protected]:0.95, Precision, and Recall, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e10\u003c/span\u003e. The PR curve represents the relationship between recall (x-axis) and precision (y-axis). The area under this curve reflects the mean Average Precision (mAP), and values closer to 1 indicate better model performance. Among the four defect categories, the mineralization category achieves the highest detection accuracy, with an mAP of 0.988, demonstrating the model's strong capability in recognizing this specific lesion type. The mAP scores for cracking and holes are 0.821 and 0.911, respectively, while the incomplete category achieves 0.946, placing its performance between that of holes and mineralization. The model shows high precision at lower recall levels, but as recall increases, the precision begins to decline due to a rise in false positives, a common trade-off in classification tasks.\u003c/p\u003e \u003cp\u003eThe Confusion Matrix (CM) is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e11\u003c/span\u003e, providing a detailed assessment of classification accuracy for each defect type. The horizontal axis denotes the ground truth labels, while the vertical axis indicates the predicted labels. The categories mineralization, holes, and incomplete exhibit strong classification accuracy. Although the model shows slightly lower performance for the cracking category, it still achieves a recognition accuracy of 84%. Overall, the EfficientNetV2-enhanced YOLOv5s model exhibits robust classification performance across all defect categories, validating its effectiveness for intelligent detection and classification of bronze disease features in cultural heritage artifacts.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Visualization results\u003c/h2\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e12\u003c/span\u003e presents the visualization interface designed for the intelligent detection of bronze diseases. The user can select a pre-trained model, such as the YOLOv5s_EfficientNetV2 model or the best-performing model, by clicking the \"Import Model\" button located in the navigation bar. After that, the user can upload an input image, namely, a bronze ware disease image\u0026mdash;via the \"Insert Image\" button. Upon clicking the \"Start Detection\" button, the system navigates to the Bronze Disease Detection page, where the uploaded image is processed, and the locations and types of bronze disease features are detected and displayed in real time. The detection results for various defect categories are shown in Figs.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e13\u003c/span\u003e through \u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e16\u003c/span\u003e, which demonstrate the model\u0026rsquo;s capability to identify and localize different types of bronze ware deterioration. Specifically: Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e13\u003c/span\u003e illustrates the recognition results for mutilation, Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e14\u003c/span\u003e shows the detection of holes, Fig.\u0026nbsp;15 presents results for whole-body mineralization, and Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e16\u003c/span\u003e highlights the model\u0026rsquo;s performance in detecting cracks. These visualizations further validate the practicality and effectiveness of the enhanced model in real-world applications, providing a user-friendly interface for cultural heritage conservation and disease diagnosis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Suggestions for the Conservation of Bronze Ware\u003c/h2\u003e \u003cp\u003eTaking the bronze artifacts housed in the Jingmen Museum as a case study, this research employed the improved YOLOv5s_EfficientNetV2 model to detect and classify surface degradation features. The enhanced model leverages compound scaling and an optimized convolutional structure to significantly improve the efficiency of multi-scale feature extraction, enabling more accurate and robust recognition of disease patterns. Furthermore, a visual interface developed with OpenCV and PyQt5 allows for intuitive visualization of the location, category, and confidence level of detected diseases. This facilitates more informed decision-making for subsequent conservation and restoration processes. Based on the diagnostic results, the following preservation and restoration strategies are proposed:\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e1. Incomplete Structures\u003c/h3\u003e\n\u003cp\u003eFor minor surface damage, missing components can be restored using welding techniques, with strict control of the welding temperature to prevent thermal damage to adjacent areas. In cases of large-scale structural loss, riveting or bonding methods are more suitable. The repaired surfaces should undergo artificial aging\u0026mdash;via chemical patination or pigment application\u0026mdash;to harmonize the color and texture with the original artifact, thereby maintaining both aesthetic integrity and historical authenticity.\u003c/p\u003e\n\u003ch3\u003e2. Holes\u003c/h3\u003e\n\u003cp\u003eBegin by using ultrasonic cleaning equipment to remove corrosion products and foreign matter from the holes. Subsequently, reinforce the fragile structures surrounding the holes using high-molecular materials such as epoxy resin or acrylic resin, which can be introduced via injection or brushing to enhance structural strength. Appropriate filling materials\u0026mdash;chosen based on the size and geometry of the holes\u0026mdash;are then embedded using welding or mechanical fixation (e.g., bolts), followed by surface filling and cosmetic restoration.\u003c/p\u003e\n\u003ch3\u003e3. Whole-body Mineralization\u003c/h3\u003e\n\u003cp\u003eFor extensively mineralized regions, corrosion suppression should be carried out using chemical inhibitors. A diluted benzotriazole (BTA) solution (typically 0.5\u0026ndash;1%) is employed, with immersion time adjusted according to the severity of mineralization. BTA molecules form a protective complex layer on the metal surface, which blocks contact with oxygen and moisture, thereby inhibiting further corrosion. After treatment, apply a silicone-based sealant (e.g., polydimethylsiloxane) via spraying or brushing to form a transparent, breathable film, which both prevents external contamination and allows internal moisture to dissipate\u0026mdash;minimizing the risk of localized corrosion due to sealing.\u003c/p\u003e\n\u003ch3\u003e4. Cracks\u003c/h3\u003e\n\u003cp\u003eNarrow cracks can be filled with low-viscosity epoxy resin and injected using a pressure syringe. The adhesive\u0026rsquo;s good fluidity ensures complete penetration and sealing, effectively preventing fissure propagation. For larger cracks (\u0026gt;\u0026thinsp;0.5 mm), the fissure edges should be cleaned using micro-mechanical tools to remove rust and debris. Then, bronze strips\u0026mdash;matching the original material\u0026mdash;are inserted and fixed via welding or bonding. The surface is then ground and aged to ensure visual and material continuity with the original artifact.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eIn this study, an enhanced YOLOv5s_EfficientNetV2 model was developed to improve the recognition of surface diseases in bronze artifacts. By integrating the lightweight and high-performance EfficientNetV2 backbone, the model demonstrates significant improvements in precision, recall, and multi-scale feature extraction. The optimal batch size was identified as 8, balancing accuracy and computational efficiency. The addition of attention mechanisms further enhanced the model's robustness in complex detection environments. A user-friendly visual interface was also developed using OpenCV and PyQt5, enabling an intuitive display of disease types, locations, and confidence levels. This tool provides practical support for cultural heritage professionals in the protection and restoration of bronze wares. Although the current dataset includes only four disease categories, the model exhibits strong scalability and adaptability. As more annotated data become available, it is expected to achieve even greater accuracy and generalization. In the long term, this method holds promise not only for broader applications in bronze conservation but also for digital preservation efforts across a wide range of cultural heritage artifacts.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eHan Wu and Jing Yang wrote the main manuscript text. All authors prepared figures and tables. All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgments\u003c/h2\u003e \u003cp\u003eThis research was supported by the Open Project of the Key Scientific Research Base of the National Cultural Heritage Administration for the Protection of Unearthed Wood Lacquerware (2021H10198, 2023H10017).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWang Z, Xi X, Li L, Zhang Z, Han Y, Wang X, Sun Z, Zhao H, Yuan N, Li H, Yan B. Chen. Tracking the progression of the simulated bronze disease\u0026mdash;a laboratory X-ray microtomography study. Molecules. 2023;28(13):4933\u0026ndash;49.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaraiva AS, Figueiredo E, \u0026Aacute;guas H, Silva RJC. Characterisation of archaeological high-tin bronze corrosion structures. Stud Conserv. 2022;67(4):222\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi J, Li L, Xie Z, Xiang J, Zhao X, Xiao Q. Ling. A comprehensive assessment method for the health status of bronzes unearthed at archaeological sites. Herit Sci. 2023;11(1):86\u0026ndash;101.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang L, Xia Q, He T, Dong W. Rational construction of a protective layer for corroded bronze based on composite corrosion inhibitors. Herit Sci. 2025;13(1):187\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu M, Yang L, Chai R. Research on multi-scale fusion method for ancient bronze ware X-ray images in NSST domain. Appl Sci. 2024;14(10):4166\u0026ndash;84.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y, Zhang S, Li W, Yue H. Application of X-Ray Technology in Detecting the Disease Structure of Bronze Cultural Relics. The 15th National Nuclear Structure Conference, Guilin, Guangxi, China, 20149(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePidaparti RM, Aghazadeh BS, Whitfield A, Bao AS. Mercier. Classification of corrosion defects in NiAl bronze through image analysis. Corros Sci. 2010;52(11):3661\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Y, Huang Y, Zhang Z, Wang Z, Liu B, Liu C, Huang C, Dong S, Pu X, Wan F, Qiao X. Qian. Plant image recognition with deep learning: A review. Comput Electron Agric. 2023;212:108072\u0026ndash;88.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe T. Image quality recognition technology based on deep learning. J Vis Commun Image Represent. 2019;65:102654\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu K, Lin K, Zhu C. Research on Chinese traditional opera costume recognition based on improved YOLOv5. Herit Sci. 2023;11(1):1\u0026ndash;14.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXia S. Research on Disease Detection Method for BronzeArtifacts based on Deep Learning. Shaanxi: Master's Thesis of Northwest University, 2022(in Chinese).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSelvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int J Comput Vision. 2020;128:336\u0026ndash;59.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXie Y, Hu W, Xie S, He L. Surface defect detection algorithm based on feature-enhanced YOLO. Cogn Comput. 2023;15(2):565\u0026ndash;79.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao Y, Wang H, Xie X, Xie Y, Yang C. An enhanced yolov5-based algorithm for metal surface defect detection. Appl Sci. 2023;13(20):11473\u0026ndash;89.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShi J, Yang J, Zhang Y. Research on steel surface defect detection based on YOLOv5 with attention mechanism. Electronics. 2022;11(22):3735\u0026ndash;49.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao Z, Bakar EBA, Razak NBA. M. N. Akhtar. Corrosion image classification method based on EfficientNetV2. Heliyon. 2024;10(17):1\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeng K, He S, Wu X, Jiang P, Huang S. Enhanced YOLOv5 for micro-defect detection on KDP crystal surfaces: a fusion of EfficientNetV2 and normalized Wasserstein distance. J Real-Time Image Proc. 2025;22(2):72\u0026ndash;83.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Z, Wu L, Li T, Shi P. A smoke detection model based on improved YOLOv5. Mathematics. 2022;10(7):1190\u0026ndash;202.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu B. Luo. An improved Yolov5 for multi-rotor UAV detection. Electronics. 2022;11(15):2330\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiang J, Kong R, Ma R, Zhang J, Bian X. Aluminum surface defect detection algorithm based on improved yolov5. Adv Theory Simulations. 2024;7(2):2300695\u0026ndash;705.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZeng L, Duan X, Pan Y, Deng M. Research on the algorithm of helmet-wearing detection based on the optimized yolov4. Visual Comput. 2023;39(5):2165\u0026ndash;75.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Y, Wang Y, Sui D, Guo M. Dense Buddha head object detection and counting YOLOv8 network based on multi-scale attention and data augmentation fusion. Herit Sci. 2025;13(1):23\u0026ndash;37.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLei X, Pan H, Huang X. A dilated CNN model for image classification. IEEE access. 2019;7:124087\u0026ndash;95.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu G, Hu Y, Chen Z, Guo J, Ni P. Lightweight object detection algorithm for robots with improved YOLOv5. Eng Appl Artif Intell. 2023;123:106217\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiong C, Hu S, Fang Z. Application of improved YOLOV5 in plate defect detection. Int J Adv Manuf Technol, 2022: 1\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLang X, Ren Z, Wan D, Zhang Y, Shu S. MR-YOLO: An improved YOLOv5 network for detecting magnetic ring surface defects. Sensors. 2022;22(24):9897\u0026ndash;913.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe K, Zhang X, Ren S, Sun J. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Trans Pattern Anal Mach Intell. 2015;37(9):1904\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi F, Xiao K, Hu Z, Zhang G. Fabric defect detection algorithm based on improved YOLOv5. Visual Comput. 2024;40(4):2309\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu S, Qi L, Qin H, Shi J, Jia J. Path aggregation network for instance segmentation. Proceedings of the IEEE conference on computer vision and pattern recognition, 2018: 8759\u0026ndash;8768.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang J, Zhang Y, Liu J, Lan Y, Zhang T. Human figure detection in Han portrait stone images via enhanced YOLO-v5. Herit Sci. 2024;12(1):1\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan M, Le Q. Efficientnet: Rethinking model scaling for convolutional neural networks. International conference on machine learning. PMLR, 2019: 6105\u0026ndash;6114.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSifre L, Mallat S. Rigid-motion scattering for texture classification. arxiv preprint arxiv:1403 1687, 2014: 1\u0026ndash;19.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan M. Q. Le. Efficientnetv2: Smaller models and faster training. International conference on machine learning, PMLR, 2021: 10096\u0026ndash;10106.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKanai S, Fujiwara Y, Yamanaka Y. S. Asachi. Sigsoftmax: Reanalysis of the softmax bottleneck. Adv Neural Inf Process Syst, 2018, 31\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHwang JS, Lee SS, Gil JW, Lee CK. Determination of optimal batch size of deep learning models with time series data. Sustainability. 2024;16(14):5936\u0026ndash;46.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Improved YOLOv5s, EfficientNetV2, Bronze Artifact Disease Recognition, Object Detection, Visual interface","lastPublishedDoi":"10.21203/rs.3.rs-6900599/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6900599/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBy leveraging the lightweight and high-performance characteristics of EfficientNetV2, this paper enhances the YOLOv5s model to accurately classify disease types in bronze wares while reducing human error. The efficiency of feature extraction across multiple scales is further improved through an optimized convolutional structure and a compound scaling method. The bronze wares housed in the Jingmen Museum serve as a case study to validate the model\u0026rsquo;s effectiveness. Photographs of the artifacts were captured using a Canon G1X Mark III digital camera and processed into a dataset comprising 1,037 images, labeled with defects such as surface flaws, holes, full-body mineralization, and cracks for training and validation. Results show that the enhanced YOLOv5s_EfficientNetV2 model significantly outperforms the original YOLOv5s model across key metrics including Precision, Recall, [email protected], and [email protected]:0.95, with respective improvements of 7.0%, 1.7%, 2.2%, and 2.6%. Furthermore, a set of visual interfaces for disease detection in bronze wares has been developed using OpenCV and PyQt5, allowing intuitive visualization of the results. The study also explores the impact of batch size on model performance, revealing that a batch size of 8 offers the best trade-off between gradient estimation accuracy and training stability, thereby enhancing recognition capability. By quantifying the type and location of deterioration, this technology offers valuable data support for conservation strategies, provides robust assistance in the protection and restoration of bronze artifacts, and demonstrates strong potential for broader applications in cultural heritage preservation.\u003c/p\u003e","manuscriptTitle":"Identification and Classification of Bronze Surface Diseases Based on Improved YOLOv5","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-23 06:27:43","doi":"10.21203/rs.3.rs-6900599/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":"f81d3770-6b39-4d3b-ab87-f2128f5f1d49","owner":[],"postedDate":"June 23rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-26T21:24:04+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-23 06:27:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6900599","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6900599","identity":"rs-6900599","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}