Photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion

Abstract

The application relates to the field of defect detection based on deep learning, and discloses a photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion, wherein: an EL image of a photovoltaic cell is acquired; a main network integrated with a photovoltaic edge information auxiliary extractor and an enhanced dilated convolution residual block in a photovoltaic defect detection model is adopted to perform feature extraction on the EL image, so that a multi-scale feature image is obtained; a multi-level context feature fusion pyramid network serving as a neck network in the photovoltaic defect detection model is adopted to perform feature fusion and enhancement on each scale feature image in the multi-scale feature image, so that a multi-scale semantic feature enhanced image is obtained; and a head network of the photovoltaic defect detection model is adopted to perform defect detection on the multi-scale semantic feature enhanced image, so that a defect type and a defect position of the photovoltaic cell are obtained. The application can improve the defect detection precision and efficiency of the photovoltaic cell.

Description

Technical Field

[0001] This application relates to the field of deep learning-based defect detection, and includes, but is not limited to, methods for detecting photovoltaic cell defects based on edge enhancement and multi-scale feature fusion. Background Technology

[0002] With the increasing global demand for sustainable energy, photovoltaic (PV) power generation, as a clean and renewable energy source, has been widely adopted globally. However, during the production and long-term operation of PV cells, surface defects such as grid breaks, cracks, black cores, coarse lines, and misalignments are prone to occur due to process defects or external environmental factors. These defects directly lead to a significant decrease in the photoelectric conversion efficiency of PV cells and may cause localized degradation of the module's electrical performance, thereby inducing early failures such as hot spot effects, localized overheating, and module failure. Because these failures lack obvious electrical characteristics, traditional monitoring methods struggle to detect them in a timely manner, leading to the continuous accumulation of potential hazards and jeopardizing the safety and operational stability of PV systems. Furthermore, according to the International Energy Agency's forecast, solar power will account for 80% of the global increase in renewable energy capacity in the coming years. Therefore, how to achieve efficient and accurate defect detection to ensure the safety and reliability of PV systems has become a crucial issue that the industry urgently needs to address. Summary of the Invention

[0003] Based on the above problems, this application provides a photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion, aiming to improve the accuracy and efficiency of photovoltaic cell defect detection.

[0004] The technical solution of this application embodiment is implemented as follows:

[0005] This application provides a photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion. The method includes: acquiring an electroluminescence (EL) image of the photovoltaic cell; using a backbone network integrating a photovoltaic edge information auxiliary extractor and an enhanced dilated convolutional residual block in a photovoltaic defect detection model to extract features from the EL image to obtain a multi-scale feature image; wherein the photovoltaic edge information auxiliary extractor introduces a hierarchical edge perception mechanism; using a multi-level context feature fusion pyramid network as the neck network in the photovoltaic defect detection model to perform feature fusion and enhancement on each scale feature image in the multi-scale feature image to obtain a multi-scale semantic feature enhanced image; and using the head network of the photovoltaic defect detection model to perform defect detection on the multi-scale semantic feature enhanced image to obtain the defect type and defect location of the photovoltaic cell.

[0006] The beneficial effects of the technical solutions provided in this application include at least the following:

[0007] This application provides a photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion. On the one hand, by introducing a photovoltaic edge information auxiliary extractor with a hierarchical edge perception mechanism and an enhanced dilated convolutional residual block, the photovoltaic defect detection model's ability to perceive weak boundaries and fine-grained defects can be enhanced, and its receptive field can be effectively expanded while improving the detection efficiency of the photovoltaic defect detection model. On the other hand, by using a multi-level context feature fusion pyramid network, the photovoltaic defect detection model's ability to adaptively fuse cross-scale features in complex backgrounds can be enhanced, and the semantic consistency of the model with respect to features can be improved. Thus, while ensuring the detection accuracy of the photovoltaic defect detection model, its real-time performance and robustness can be significantly improved, thereby improving the accuracy and efficiency of the photovoltaic defect detection model in detecting defects in photovoltaic cells. Attached Figure Description

[0008] Figure 1 A schematic flowchart illustrating a photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion provided in this application embodiment;

[0009] Figure 2 This is a schematic diagram of the structure of EDRB-MF provided in an embodiment of this application;

[0010] Figure 3 A schematic diagram showing the connection of three cascaded edge feature transmission fusion modules in the PV-EIAE provided in this embodiment of the application;

[0011] Figure 4 A schematic diagram of the MCF-FPN framework provided in the embodiments of this application;

[0012] Figure 5 A schematic diagram of the MEE-YOLO framework provided for an embodiment of this application;

[0013] Figure 6 This is a comparison chart of the detection results of the photovoltaic defect detection model and the YOLOv11n model disclosed in the embodiments of this application under different defect types. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0016] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0017] In related technologies, accurate detection of surface defects in photovoltaic cells is crucial for ensuring the stable operation of solar energy systems. However, defect detection based on EL images faces the dual challenges of complex lattice texture backgrounds and multi-scale defects, leading to limitations in feature extraction and low efficiency in semantic fusion during detection.

[0018] Traditional photovoltaic cell defect detection methods mainly include manual inspection, physical inspection, and machine vision-based inspection methods. Manual inspection relies heavily on worker experience, is easily affected by environmental interference, and suffers from low efficiency and high cost, leading to its gradual phasing out. Physical inspection (e.g., mechanical impact testing, current / voltage curve measurement, resonant ultrasonic vibration, etc.) is mostly used for online diagnostics during wafer and cell manufacturing. While offering high accuracy, it often involves significant human intervention in practical applications, resulting in complex procedures and hindering automation. In contrast, traditional machine vision-based inspection methods (including image domain analysis and transform domain analysis) offer advantages such as high speed, low cost, and high accuracy. However, these methods typically rely on manually designed feature extraction algorithms, requiring algorithm customization for different types of defects. They also lack generalization ability when dealing with complex or irregular defects such as microcracks and hidden cracks.

[0019] Compared to traditional machine vision-based detection methods, deep learning models trained on large-scale datasets exhibit stronger feature representation capabilities and generalization adaptability, making them a research hotspot in photovoltaic defect detection. Based on this, this application focuses on the EL image defect detection task of photovoltaic modules. Since the You Only Look Once (YOLO) series, as a typical single-stage object detection framework, achieves a better trade-off between speed and accuracy, it is particularly suitable for industrial applications that demand both detection accuracy and real-time performance. YOLOv11, with its efficient end-to-end structure and excellent balance between speed and accuracy, has been selected as the base model for the photovoltaic defect detection model provided in this application. However, when YOLOv11, as a general object detection framework, is directly applied to the EL image defect detection task of photovoltaic cells, it still faces the following challenges: First, the complex lattice background causes the defect edges to be blurred, and YOLOv11 has insufficient perception ability in fine-grained boundary modeling; second, its ability to model long-range semantic relationships is limited, affecting the identification and localization of defect targets in complex backgrounds; and third, there is significant information loss in the multi-scale feature fusion process, which limits the localization accuracy of YOLOv11 for defects of different sizes and shapes.

[0020] Based on the above description, embodiments of this application provide a photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion, referencing... Figure 1 As shown, the method includes the following steps:

[0021] Step 101: Obtain the EL image of the photovoltaic cell.

[0022] In some embodiments, photovoltaic cells are the core unit of photovoltaic modules, used to convert solar energy into electrical energy. An EL image of a photovoltaic cell is a two-dimensional grayscale image obtained by applying a forward bias voltage to the photovoltaic cell to excite its internal electroluminescence response, containing typical defects such as cracks, broken metal grids, and black cores.

[0023] It should be noted that compared with infrared or optical images, EL images have advantages such as high resolution, clear details, and stable response, and have been widely used in quality inspection during the manufacturing and operation and maintenance of photovoltaic modules.

[0024] Here, the number of EL images corresponding to the photovoltaic cells can be 1, 2, 3 or more, and this application does not impose any limitation on this.

[0025] Step 102: Using the backbone network of the photovoltaic defect detection model, which integrates a photovoltaic edge information auxiliary extractor and an enhanced dilated convolutional residual block, feature extraction is performed on the EL image to obtain a multi-scale feature image.

[0026] Among them, the photovoltaic edge information auxiliary extractor introduces a hierarchical edge perception mechanism.

[0027] In some embodiments, the photovoltaic defect detection model can be a model generated by optimizing and adjusting YOLOv11 (see the model structure provided by the Ultralytics team) as the basic framework.

[0028] In some embodiments, the construction process of the photovoltaic defect detection model can be achieved through the following steps A1 to A4:

[0029] Step A1: Obtain the initial detection model built by YOLOv11.

[0030] Step A2: Integrate a photovoltaic edge information auxiliary extractor into the backbone network of the initial detection model, and simultaneously replace the C3K2 module used for feature extraction in the initial detection model with an enhanced dilated convolutional residual block to obtain an intermediate detection model.

[0031] In some embodiments, the initial detection model built with YOLOv11 mainly includes a backbone network, a neck network, and a head network. Here, the backbone network and the neck network can be improved (adjusted) respectively.

[0032] In some embodiments, edge features, as high-frequency structural information in an image, can provide clear target contours and positioning clues, especially suitable for background scenes with strong lattice texture interference and weak contrast in EL images. To address this, this application embeds a Photovoltaic-Edge Information Auxiliary Extractor (PV-EIAE) with a hierarchical edge perception mechanism into the backbone network of the initial detection model. This enables the dynamic extraction of multi-scale edge information from the shallow features of the EL image and the construction of an explicit structural information transmission path. Furthermore, the edge feature transmission module in PV-EIAE is used to gradually inject edge semantics, i.e., multi-scale edge information, into the deep features of the EL image (obtained by extracting relevant features of the EL image using the enhanced dilated convolution residual block in PV-EIAE). This achieves structure-guided semantic modeling, effectively preserving the edge contour information of the EL image, thereby improving the response consistency to blurred defect areas in the EL image.

[0033] In some embodiments, the PV-EIAE may include four parts: 1. A MultiScale Edge Information Generator (MSEIG) for dynamically extracting multiscale edge information from the shallow features of the EL image; 2. An Enhanced Dilated Residual Block for Multi-Scale Feature Extraction (EDRB-MF) for extracting high-level semantic representations, i.e., the deep visual features of the EL image, from the EL image; 3. An Edge Information Transfer (EITransfer) module for selecting features of corresponding scales from the multiscale edge information and transferring them to the Edge Information Fusion (EIFusion) module; 4. EIFusion for hierarchically fusing the multiscale edge information dynamically extracted from the shallow features of the EL image with the high-level semantic representations, i.e., the deep visual features of the EL image.

[0034] Here, in PV-EIAE, apart from MSEIG, EDRB-MF or a convolutional encoder block (CE-Block) integrated with EDRB-MF, EIFusion, and EITransfer can be connected in a cascaded manner, and the specific connection method can be determined according to actual needs.

[0035] In some embodiments of this application, EDRB-MF is a module proposed to address the problem that traditional convolutional structures, when operating within a fixed receptive field, struggle to simultaneously capture both large-scale contextual information and local detail representation, leading to potential feature omissions or blurred boundaries in detection results. It is based on multi-scale receptive field modeling, integrating a dilation-wise residual (DWR) module and a dilated reparameterized block (DRB) module to construct a feature representation unit (Darknet ResidualBlock Bottleneck, DRR Bottleneck) that combines contextual modeling capabilities with structural compactness, enabling the extraction of deep visual features from the input feature image. Its structure is as follows: Figure 2 As shown. For the feature image 201 input to EDRB-MF, its feature image channels have a cardinality of c, and it is processed sequentially through region residualization and semantic residualization. Furthermore, this EDRB-MF also involves operations such as standard convolution, depthwise convolution, and batch normalization. Figure 2The dilated reparameterization module shown can be dilated reparameterized convolution with dilation rates of 5 and 3, respectively. The "+" inside the circle indicates element-wise addition.

[0036] It should be noted that the C3K2 module used for feature extraction in the initial detection model is a composite module that integrates downsampling and Cross Stage Partial Network (CSPNet). It is a common module in YOLOv11. The C3 representation in the C3K2 module contains a CSPNet with 3 convolutions, and the kernel size used in the K2 representation is 2×2.

[0037] Step A3: Change the neck network in the intermediate detection model to a multi-level context feature fusion pyramid network to obtain an improved initial detection model.

[0038] In some embodiments, the Multilevel Contextual Feature Fusion Pyramid Network (MCF-FPN) can be collaboratively designed by a Feature Selection Module (FSM), a Content-Aware ReAssembly of Feature Elements (CARAFE), and an Adaptive Feature Fusion (AFF) module to achieve the construction of an MCF-FPN with a multi-scale fusion mechanism that has semantic awareness, spatial reconstruction, semantic hierarchical alignment, and structural restoration capabilities.

[0039] In some embodiments, the initial detection model built by YOLOv11 is improved by integrating PV-EIAE to enhance the initial detection model's ability to perceive the edges of subtle defects in complex backgrounds; and based on EDRB-MF, a reparameterized multi-branch expansion structure is used to fuse different receptive fields to enhance the initial detection model's ability to model semantic context, thereby improving the initial detection model's ability to represent multi-scale features; and with the help of MCF-FPN, a channel attention mechanism and a content-aware upsampling operator are introduced to further enhance the initial detection model's adaptive fusion ability and semantic consistency of cross-scale features in complex backgrounds, thereby improving the accuracy and robustness of the improved initial detection model in the defect detection process.

[0040] Step A4: Using the photovoltaic cell defect anomaly detection dataset as the training set, train the improved initial detection model by minimizing the loss function to obtain the photovoltaic defect detection model.

[0041] In some embodiments, the Photovoltaic Electroluminescence-Anomaly Detection dataset (PVEL-AD), which is used as a benchmark dataset for methods of detecting anomalies in photovoltaic cells, can be used as a photovoltaic cell defect anomaly detection dataset. PVEL-AD contains 36,543 near-infrared images with various internal defects and heterogeneous backgrounds. Among them, there is one class of images without anomalies and images with 12 different categories of anomalies, such as: cracks, star cracks, finger lines, fragments, black cores, horizontal dislocations, vertical dislocations, thick lines, print errors, corners, scratches, short circuits, etc.

[0042] In some embodiments, PVEL-AD is used as the training set, and the improved initial detection model is iteratively trained by minimizing the loss function until a photovoltaic defect detection model whose output loss meets preset conditions is obtained. In this way, a photovoltaic defect detection model with improved detection accuracy and robustness can be obtained, thus providing a computational foundation for improving the accuracy and efficiency of subsequent defect detection of photovoltaic cells.

[0043] In some embodiments, the initial feature extractor integrated with EDRB-MF in the backbone network of the photovoltaic defect detection model is first used to extract the original visual features of the EL image to obtain the original visual feature image. Then, the PV-EIAE (which may also include multiple cascaded EDRB-MFs) in the backbone network with a hierarchical edge perception mechanism is used to extract multi-scale edge features and fuse cross-level features in the original visual feature image to obtain the multi-scale feature image corresponding to the EL image.

[0044] Step 103: Using the multi-level context feature fusion pyramid network as the neck network in the photovoltaic defect detection model, feature fusion and enhancement are performed on each scale feature image in the multi-scale feature image to obtain a multi-scale semantic feature enhanced image.

[0045] In some embodiments, MCF-FPN can be used to perform feature fusion and enhancement on each scale feature image in the multi-scale feature image, resulting in a multi-scale semantic feature enhanced image composed of semantic feature enhanced images corresponding to each scale feature image. Simultaneously, MCF-FPN can utilize its internal channel attention guidance and upsampling operators, such as CARAFE, to achieve adaptive fusion of cross-scale semantic features in the multi-scale feature image. This can mitigate feature degradation during the upsampling process and significantly improve the detection accuracy of the photovoltaic defect detection model for defects of different sizes and shapes.

[0046] Step 104: Using the head network of the photovoltaic defect detection model, perform defect detection on the multi-scale semantic feature enhanced image to obtain the defect type and defect location of the photovoltaic cell.

[0047] In some embodiments, the head network of the photovoltaic defect detection model can be used to perform defect detection on the multi-scale semantic feature enhanced image to obtain the defect type (e.g., black core, crack, scratch, fragment, etc.) and its corresponding location for each defect of the photovoltaic cell, i.e., the defect location (middle position, edge position, etc.).

[0048] In other words, this application's embodiments propose a photovoltaic cell defect detection method based on YOLOv11, integrating edge-aware enhancement and multi-scale semantic fusion mechanisms. It also provides a framework specifically designed for photovoltaic panel defect detection based on improved YOLOv11 (Multi-Exposure Ensemble-YOLO, MEE-YOLO). Firstly, to address the insufficient utilization of edge information in traditional frameworks, PV-EIAE is designed in the backbone network of the photovoltaic defect detection model to enhance its edge-awareness of subtle defects in complex backgrounds. Secondly, EDRB-MF is proposed to replace the original C3K2 module, enabling the fusion of different receptive fields using a reparameterized multi-branch expansion structure, enhancing semantic context modeling capabilities, and improving the photovoltaic defect detection model's representation of multi-scale features. Finally, MCF-FPN is introduced into the photovoltaic defect detection model, further enhancing the adaptive fusion capability and semantic consistency of cross-scale features in complex backgrounds through channel attention mechanisms and content-aware upsampling operators. This allows for the improvement of both the real-time performance and robustness of photovoltaic defect detection while ensuring accuracy.

[0049] The photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion provided in this application has several advantages. First, by introducing a photovoltaic edge information auxiliary extractor with a hierarchical edge perception mechanism and an enhanced dilated convolutional residual block, the photovoltaic defect detection model can not only enhance its ability to perceive weak boundaries and fine-grained defects, but also effectively expand its receptive field while improving its detection efficiency. Second, by leveraging a multi-level contextual feature fusion pyramid network, the photovoltaic defect detection model can enhance its adaptive fusion capability and semantic consistency across scale features in complex backgrounds. Thus, while ensuring the accuracy of photovoltaic defect detection, the real-time performance and robustness of photovoltaic defect detection can be significantly improved, thereby enhancing the accuracy and efficiency of photovoltaic cell defect detection.

[0050] In some embodiments, step 102 provided in the above embodiments can be implemented by steps 1021 and 1022. Figure 1 (not shown):

[0051] Step 1021: Use the initial feature extractor with enhanced dilated convolution residual blocks integrated in the backbone network to extract the original visual features of the EL image and obtain the original visual feature image.

[0052] In some embodiments, the initial feature extractor integrated with EDRB-MF can extract raw visual features from the EL image through a two-stage process of region residualization and semantic residualization, so as to obtain a raw visual feature image with enhanced integration capability of multi-scale semantic context and improved sensitivity to structural response of key regions.

[0053] In some embodiments, step 1021 can be implemented by steps 10211 and 10212. Figure 1 (not shown):

[0054] Step 10211: Using multiple parallel convolutional branches with different dilation rates in the enhanced dilated convolutional residual block, multi-scale contextual features are extracted in parallel from the EL image to obtain multiple sets of initial features.

[0055] Step 10212: Using the semantic residual fusion module in the enhanced dilated convolution residual block, multiple sets of initial features are integrated to obtain integrated features. The integrated features are then residually connected with the EL image to obtain the original visual feature image with high-level semantic information.

[0056] In some embodiments, to ensure the context-awareness of EDRB-MF and maintain semantic consistency, EDRB-MF employs a two-stage process of "region residualization-semantic residualization." First, multiple parallel convolutional branches with different dilation rates in EDRB-MF are used to extract multi-scale contextual information from the EL image in parallel, explicitly expanding the receptive field. This helps capture multiple sets of initial features corresponding to the long-range dependencies between microcracks and large-scale defects in the EL image. Then, the semantic residual fusion module (path) in EDRB-MF is used to uniformly align (integrate) the extracted multiple sets of initial features, i.e., multi-scale features. The integrated features are then residually connected to the EL image to obtain an original visual feature image with high-level semantic information. This alleviates semantic offset and boundary errors caused by scale differences and spatial resolution inconsistencies, thereby improving the backbone network of the photovoltaic defect detection model's ability to locate defect regions in the EL image and maintain structural integrity.

[0057] Following the description above, EDRB-MF, with multi-scale receptive field modeling as its core, integrates DWR and DRB to construct a DRR Bottleneck that combines contextual modeling capabilities with structural compactness. (See also: [link to relevant documentation]). Figure 2 Among them, DRR Bottleneck can introduce semantic information at different scales through multi-branch expandable residual convolutions, enhancing the ability to model long-range dependencies. Meanwhile, DRB can introduce redundant paths to improve expressive power during the training phase of the photovoltaic defect detection model, and achieve path fusion through expansion and reparameterization during the inference phase, significantly reducing the computational cost of the photovoltaic defect detection model. Specifically, DRB uses a multi-branch expandable convolutional structure during the training phase to fuse information from different receptive fields, improving nonlinear representation and local structural response. During the inference phase, the branches of DRB are integrated into a single convolutional kernel through reparameterization, thereby significantly compressing the network structure complexity while maintaining feature expressive power.

[0058] Here, EDRB-MF enhances the photovoltaic defect detection model's ability to integrate multi-scale semantic context in EL images and its sensitivity to structural responses in key regions of EL images by fusing multi-scale expanded receptive field, semantic residual fusion path, and reparameterization mechanism. While ensuring controllable parameter scale, EDRB-MF achieves synergistic optimization of receptive field coverage, region response focusing, and inference efficiency, making it suitable for complex defect detection scenarios in EL images that simultaneously contain large-sized dark flaws and micro-cracks. Compared to the C3K2 module, EDRB-MF achieves a better trade-off between feature representation capability and deployment efficiency.

[0059] Step 1022: Using the photovoltaic edge information auxiliary extractor in the backbone network, multi-scale edge feature extraction and cross-level feature fusion are performed on the original visual feature image to obtain a multi-scale feature image.

[0060] In some embodiments, a PV-EIAE with a hierarchical edge-aware mechanism is introduced. First, the internal MSEIG is used to extract multi-scale edge features from the original visual feature image, resulting in a multi-scale edge feature image set. Then, multiple (e.g., three, as illustrated in the following embodiments) cascaded edge feature transfer and fusion modules (each module may consist of an EITransfer, a CE-Block carrying EDRB-MF, and an EIFusion) are used to progressively fuse the multi-scale edge feature image set with the original visual feature image, yielding a multi-scale feature image. Here, when the number of modules corresponding to the multiple cascaded edge feature transfer and fusion modules is three, the number of images included in the multi-scale edge feature image set is also three.

[0061] In some embodiments, step 1022 can be implemented by the following steps 10221 and 10222. Figure 1 (not shown in the image)

[0062] Step 10221: Use the multi-scale edge information generator in the photovoltaic edge information auxiliary extractor to extract multi-scale edge features from the original visual feature image to obtain a set of multi-scale edge feature images.

[0063] In some embodiments, the process of MSEIG operating on the original visual feature image may sequentially include: Sobel convolution → max pooling → 1x1 convolution. The specific operations can be implemented through steps B1 to B3:

[0064] Step B1: Using the Sobel operator in the multi-scale edge information generator, determine the horizontal gradient component in the horizontal direction and the vertical gradient component in the vertical direction of the original visual feature image, and construct the initial edge feature image based on the horizontal gradient component and the vertical gradient component.

[0065] In some embodiments, the initial extraction of edge information in PV-EIAE is performed by MSEIG. As shown in the following formula (1), given the input: the original visual feature image, the Sobel operator is used to calculate the horizontal gradient component in the horizontal direction and the vertical gradient component in the vertical direction of the original visual feature image respectively. After summing the squares of the horizontal gradient component and the squares of the vertical gradient component, the square root operation is performed to obtain the initial edge feature image.

[0066] Step B2: Use the max pooling layer in the multi-scale edge information generator to construct the pyramid-shaped edge representation feature image corresponding to the initial edge feature image.

[0067] Step B3: Using the convolutional layer in the multi-scale edge information generator, channel compression and feature alignment are performed on the pyramid-shaped edge representation feature image to obtain a set of multi-scale edge feature images.

[0068] In some embodiments, considering the significant differences in scale and morphology among different types of defects, inspired by scale-space theory, max-pooling is introduced into MSEIG to construct a pyramid-shaped edge representation structure to extract edge features under different receptive fields. This operation not only helps capture the edge cues of local details and global contours in the initial edge feature image, but also effectively suppresses high-frequency noise interference caused by lattice textures in the initial edge feature image. Subsequently, 1×1 convolutions in MSEIG are used for channel compression and feature alignment to form a unified set of multi-scale edge feature images, facilitating subsequent cross-scale transmission and fusion. Thus, by fusing the Sobel operator, max-pooling layers, and convolutional layers into MSEIG, multi-scale edge information is dynamically extracted from the shallow structural features of the original visual feature image, providing parameter support for subsequent fusion of edge and semantic features.

[0069] Step 10222: Using multiple cascaded edge feature transmission and fusion modules in the photovoltaic edge information auxiliary extractor, the multi-scale edge feature image set and the original visual feature image are fused step by step to obtain the multi-scale feature image.

[0070] The number of modules corresponding to multiple cascaded edge feature transmission and fusion modules is the same as the number of images included in the multi-scale edge feature image set.

[0071] In some embodiments, three cascaded edge feature transmission and fusion modules within the PV-EIAE are used to progressively fuse a set of multi-scale edge feature images with the original visual feature image to obtain a multi-scale feature image. For example... Figure 3 As shown, each edge feature transmission and fusion module 301 includes: an EITransfer, a CE-Block carrying EDRB-MF, and an EIFusion. The CE-Block carrying EDRB-MF in the first-level edge feature transmission and fusion module is used to connect to the initial feature extractor integrating EDRB-MF in the backbone network of the photovoltaic defect detection model. Simultaneously... Figure 3 Figure 303 shows a CE-Block carrying EDRB-MF, which may also include a: convolutional layer-batch normalization layer-activation function (Conv-BN-SiLU, CBS).

[0072] Correspondingly, step 10222 above can be implemented by following steps C1 to C4:

[0073] Step C1: For each level of edge feature transmission and fusion module, the edge feature transmission module in the edge feature transmission and fusion module is used to select the scale edge feature image corresponding to the cascade to which the edge feature transmission and fusion module belongs from the multi-scale edge feature image set.

[0074] Step C2: Use the enhanced dilated convolution residual block in the edge feature transfer fusion module to extract features from the intermediate feature image, and obtain the efficient feature representation image of the edge feature transfer fusion module.

[0075] In the case where the edge feature transmission and fusion module is the first-level edge feature transmission and fusion module among multiple cascaded edge feature transmission and fusion modules, the original visual feature image is the intermediate feature image; if the edge feature transmission and fusion module is not the first-level edge feature transmission and fusion module among multiple cascaded edge feature transmission and fusion modules, the intermediate feature image is the output feature image corresponding to the previous level edge feature transmission and fusion module.

[0076] In some embodiments, the CE-Block carrying EDRB-MF in the edge feature transfer fusion module is used to extract multi-scale contextual features from the intermediate feature image in parallel through multiple parallel convolution branches with different dilation rates, resulting in multiple sets of features. The semantic residual fusion module in EDRB-MF is then used to integrate the multiple sets of features, and the integrated features are residually connected with the intermediate feature image to obtain an efficient feature representation image with high-level semantic information.

[0077] Here, the specific operation of feature extraction of the intermediate feature image using the CE-Block carrying EDRB-MF in the edge feature transfer fusion module is the same as the operation in steps 10211 and 10212 above. For a detailed description, please refer to steps 10211 and 10212, which will not be repeated here. Similarly, in PV-EIAE, the corresponding feature extraction operation is also performed using the CE-Block carrying EDRB-MF.

[0078] In some embodiments, continue to refer to Figure 3As shown, in PV-EIAE, the CE-Block and MSEIG carrying EDRB-MF in the first-level edge feature transfer and fusion module are both connected to the initial feature extractor integrating EDRB-MF in the backbone network of the photovoltaic defect detection model. Furthermore, the EITransfer in each level's edge feature transfer and fusion module is used to select the corresponding scale edge feature image from the multi-scale edge feature image set output by MSEIG. The EIFusion in each level's edge feature transfer and fusion module receives the efficient feature representation image output by its internal CE-Block and the scale edge feature image output by EITransfer, respectively. The CE-Block carrying EDRB-MF in the current level's edge feature transfer and fusion module is connected to the EIFusion in the previous level's edge feature transfer and fusion module.

[0079] Here, in each level of the edge feature transfer and fusion module, EITransfer selects edge images of key scales from the multi-scale edge feature set. Employing feature embedding and spatial alignment strategies, it embeds shallow edge structure information into high-level semantic features, forming a structure-guided semantic modeling mechanism. Compared to traditional skip connections or residual fusion methods, EITransfer places greater emphasis on maintaining the continuity of edge structures, which helps enhance the boundary response and target consistency of small defect regions in complex backgrounds.

[0080] Step C3: Use the convolutional edge fusion module in the edge feature transfer fusion module to perform feature fusion on the scale edge feature image and the efficient feature representation image to obtain the output feature image corresponding to the edge feature transfer fusion module.

[0081] In some embodiments, step C3 can be implemented by the following steps C31 and C32:

[0082] Step C31: Use the stitching layer in the convolutional edge fusion module to stitch together the scale edge feature image and the efficient feature representation image to obtain the stitched feature image.

[0083] Step C32: Using the convolutional layer in the convolutional edge fusion module, channel compression, local perception enhancement, and fusion reconstruction are performed sequentially on the stitched feature image to obtain the output feature image corresponding to the edge feature transmission fusion module.

[0084] In some embodiments, continue to refer to Figure 3 As shown in 302, the EIFusion in the first-level edge feature transmission and fusion module receives the output from the backbone network. (Taking the first-level edge feature transmission and fusion module as an example,) (Original visual feature image) and edge features at corresponding scales The process begins by concatenating the data along the channel dimension to preserve the respective feature representation advantages. Then, 1×1, 3×3, and 1×1 convolution operations are sequentially applied to achieve channel compression, local perception enhancement, and fusion reconstruction, resulting in the output feature image. That is, the convolutional layer of this EIFusion ( Figure 3 The 302 layer (in the model) includes: a first 1×1 convolutional layer to reduce dimensionality and improve computational efficiency; a second 3×3 convolutional layer to capture the spatial relationship between edges and semantics; and a third 1×1 convolutional layer to integrate information and complete the output mapping. This EIFusion has stronger learnability and nonlinear modeling capabilities in its structural design, effectively enhancing the fusion quality between edge structures and semantic features, and strengthening the model's discriminative ability in complex boundary regions.

[0085] Step C4: Determine the multi-scale feature image composed of the output feature images corresponding to the edge feature transmission and fusion modules at each level.

[0086] In some embodiments, the output feature images corresponding to the edge feature transmission and fusion modules at each level can be integrated to form a multi-scale feature image that needs to be input into the neck network that makes up the photovoltaic defect detection model.

[0087] In this way, the photovoltaic defect detection model, leveraging a backbone network integrating PV-EIAE and EDRB-MF, implements a hierarchical edge-aware modeling process. MSEIG in PV-EIAE dynamically extracts multi-scale edge information from shallow features of the EL image, injects it step-by-step into high-level semantic representation via EITransfer, and finally fuses it with deep visual features, i.e., the high-level semantic representation of the EL image, through EIFusion. This hierarchical design allows the photovoltaic defect detection model to continuously retain edge cues at each stage of semantic abstraction, thereby improving the consistency and completeness of defect boundary representation and effectively enhancing the backbone network's ability to perceive defect boundaries in EL images.

[0088] In some embodiments, step 103 above can be implemented by steps 1031 and 1032. Figure 1 (not shown in the image)

[0089] Step 1031: Using the feature selection module in the multi-level context feature fusion pyramid network, feature recalibration is performed on each scale feature image in the multi-scale feature image to obtain the multi-scale enhanced feature image.

[0090] The MCF-FPN disclosed in the embodiments of this application, such as Figure 4As shown, it mainly consists of an FSM with a Channel And Context Aware Module (CACAM) and an AFF with CACAM, thus enabling MCF-FPN to integrate a multi-scale fusion mechanism with semantic awareness, spatial reconstruction, semantic hierarchical alignment, and structural restoration capabilities. Here, using the FSM in MCF-FPN, feature recalibration is performed on each scale feature image (S3, S4, and S5) in the multi-scale feature images, i.e., the feature selection stage. To enhance the selective representation capability between multi-scale semantics, MCF-FPN also integrates CACAM. The goal of CACAM is to dynamically evaluate the semantic contribution of each scale feature image on different channels, thereby achieving efficient information filtering and compression.

[0091] It should be noted that, Figure 4 The Attentional Feature Fusion Module with Multi-Aspect (AFF-MA) in the text is called AFF.

[0092] In some embodiments, step 1031 above can be implemented by steps 10311 to 10314. Figure 1 (not shown in the image)

[0093] Step 10311: Based on the feature image at each scale, the global average pooling layer and the global maximum pooling layer in the feature selection module are used to determine the global average channel-level features and the global maximum channel-level features of the scale feature image, respectively.

[0094] Step 10312: Using the fully connected layer and nonlinear activation function in the feature selection module, the global average channel-level features and the global maximum channel-level features are fused to obtain the channel attention image of the scale feature image.

[0095] Step 10313: Using the adaptive attention module in the feature selection module, the scale feature image is weighted based on the channel attention image of the scale feature image to obtain the enhanced feature image corresponding to each scale feature image.

[0096] Step 10314: Determine the multi-scale enhanced feature image composed of the enhanced feature images corresponding to each scale feature image.

[0097] In some embodiments, continue to refer to Figure 4 As shown, for each scale feature image in a multi-scale feature image To capture global response distribution and extreme value information, channel-level statistical features can be extracted first using global average pooling (Avg Pool) and global max pooling (Max Pool), respectively. Then, a fused-aware channel attention image is formed by superimposing a non-linear activation (Sigmoid) on two fully connected layers with shared parameters. The channel-selective enhancement is achieved by applying the channel-selective enhancement as a weight to the original feature image. This mechanism draws on the channel modeling idea of ​​the Squeeze and Excitation Network (SENet), which improves the representational differences between channel features at different scales, helps reduce redundancy and enhance semantic selection ability. The corresponding formula is shown in formula (1) below:

[0098] Formula (1);

[0099] in, and These are parameters that need optimization; and These represent the results of global average pooling and global max pooling, respectively. Represents the ReLU activation function. This indicates the Sigmod activation function.

[0100] Step 1032: Using the adaptive feature fusion module in the multi-level context feature fusion pyramid network, multi-level semantic feature fusion is performed on each scale enhanced feature image in the multi-scale enhanced feature image to obtain a multi-scale semantic feature enhanced image.

[0101] In some embodiments, firstly, based on each scale-enhanced feature image, a content-aware feature reconstruction operator in the adaptive feature fusion module is used to determine the attention weight of the high-level enhanced feature image associated with the scale-enhanced feature image in the multi-scale enhanced feature image. Secondly, the attention weight is used to weight the scale-enhanced feature image to obtain a weighted feature image of the scale-enhanced feature image. Thirdly, the high-level enhanced feature image is used to enhance the weighted feature image of the scale-enhanced feature image element by element to obtain an intermediate enhanced feature image of the scale-enhanced feature image. Then, the kernel prediction image of the high-level enhanced feature image is used to fuse the intermediate enhanced feature image of the scale-enhanced feature image to obtain a semantic fusion feature image of the scale-enhanced feature image. The kernel prediction image of the high-level enhanced feature image is generated by upsampling the high-level enhanced feature image using a content-aware feature reconstruction operator. Finally, a multi-scale semantic feature enhancement image composed of the semantic fusion feature images of each scale-enhanced feature image is determined.

[0102] Here, to further enhance the structural restoration capability of deep semantic features during the upsampling process, MCF-FPN employs CARAFE to replace traditional upsampling operations such as bilinear interpolation and transposed convolution, achieving adaptive reconstruction and fine fusion of high-resolution features. CARAFE adaptively generates sampling kernels based on the input feature content, enabling structure-aware reconstruction of high-resolution features. In its implementation, CARAFE first uses lightweight convolution to generate a position-related reconstructed weight image of the input features, and then completes the non-uniform restoration of spatial details through local aggregation operations. This CARAFE possesses stronger semantic information perception capabilities, improving the spatial reconstruction quality in blurred edge regions and complex texture backgrounds without introducing additional parameters, exhibiting superior structure preservation capabilities compared to transposed convolution, which is particularly crucial for restoring structural details in EL images.

[0103] Here, CARAFE is essentially a content-aware feature upsampling operator, comprising two steps: an upsampling kernel prediction step and a content-aware reconstruction step. In the upsampling kernel prediction step, given the scale-enhanced feature image in this embodiment, CARAFE generates a new feature map. Specifically, the operations include: first, dimensionality reduction of the scale-enhanced feature image to obtain a dimensionality-reduced feature image; second, spatial expansion of the dimensionality-reduced feature image to obtain an expanded feature image; then, upsampling of a pixel with a channel width and height of 1 from the expanded feature image to obtain the upsampling kernel corresponding to that pixel; thus, a corresponding upsampling kernel can be predicted for each pixel of the scale-enhanced feature image; finally, for the scale-enhanced feature image, integrating the upsampling kernels corresponding to all its pixels regenerates a kernel prediction map, resulting in a new feature map.

[0104] Meanwhile, in the content-aware reconstruction step of CARAFE, for each pixel in the new feature map, the position of the pixel on the scale-enhanced feature image can be found according to the upsampling ratio. Then, based on the position, all feature vectors in the corresponding neighborhood are selected. In the upsampling kernel prediction step of CARAFE, the upsampling kernel corresponding to the pixel is multiplied element-wise and summed to obtain the target value corresponding to the pixel, that is, the perceptual reconstruction value. By analogy, it can be determined that the semantic fusion feature image of the scale-enhanced feature image is obtained by using CARAFE.

[0105] In some embodiments, taking the multi-scale enhanced feature images as including: image S1, image S2, and image S3, where image S1 is a low-level high-resolution scale image (containing rich detail features), image S2 is a mid-level medium-resolution scale image (containing intermediate features), and image S3 is a high-level low-resolution scale image (having the largest receptive field among images S1, S2, and S3, containing high-level semantic information), the high-level enhanced feature image associated with image S1 is image S3; the high-level enhanced feature image associated with image S2 is image S3; and the high-level enhanced feature image associated with image S3 is image S1. Correspondingly, for image S1, firstly, CARAFE is used to determine the attention weight of image S3 on image S1; secondly, image S1 is weighted using this attention weight to obtain the weighted feature image of image S1; then, image S3 is used to enhance the weighted feature image of image S1 element by element to obtain the intermediate enhanced feature image of image S1; finally, the kernel prediction image of image S3 (obtained by upsampling image S3 using CARAFE) is used to fuse the intermediate enhanced feature image of image S1 to obtain the semantic fusion feature image of image S1. Similarly, the semantic fusion feature images of images S2 and S3 can be obtained, thus obtaining a multi-scale semantic feature enhancement image corresponding to the multi-scale enhanced feature images of images S1, S2, and S3.

[0106] Thus, in the semantic fusion stage, MCF-FPN introduces, for example... Figure 4 The AFF shown achieves structural consistency alignment between multi-level semantic features and spatial details, ultimately outputting multi-scale semantic feature-enhanced images: P3, P4, and P5. Traditional methods often employ feature concatenation or linear fusion strategies, which easily lead to channel redundancy and feature shift issues. Given high-level and low-level features, the AFF generates fusion attention weights through high-level semantic features, and performs element-wise addition and multiplication enhancement on the low-level features, forming a semantically driven multi-scale semantic fusion. This approach not only effectively preserves shallow spatial boundary information but also avoids information redundancy caused by concatenation, and improves the consistency and accuracy of fusion through attention modeling.

[0107] In this embodiment, to compensate for the shortcomings of existing methods in the structural perception dimension and to improve YOLO's semantic understanding ability in complex backgrounds, an edge information extraction mechanism with cross-scale guidance capability and a context enhancement structure with structural compactness and sufficient expression are provided. This enhances the photovoltaic defect detection model's ability to perceive weak defect boundaries, thereby compensating for the shortcomings of existing methods in the structural perception dimension while also considering semantic modeling capability and inference efficiency. Specifically, a hierarchical edge perception mechanism, PV-EIAE, is integrated into the backbone network of the photovoltaic defect detection model. This mechanism embeds multi-scale edge feature generation and cross-level fusion mechanisms into the backbone network, effectively improving the photovoltaic defect detection model's ability to perceive weak boundaries and fine-grained defects in EL images. This, in turn, enhances the robustness of the photovoltaic defect detection model in detecting photovoltaic cells under complex background interference. Furthermore, by leveraging EDRB-MF and integrating multi-rate convolutional branches and reparameterization mechanisms, an efficient feature representation structure with hierarchical semantic modeling capabilities is constructed. While maintaining controllable computational overhead, this EDRB-MF significantly expands the receptive field coverage of the photovoltaic defect detection model, effectively suppressing background noise interference, thereby enhancing the semantic representation ability of extracted mid-to-high-level features for complex defect regions. Simultaneously, a multi-scale feature fusion framework with context-guided and structurally adaptive mechanisms, namely MCF-FPN, is designed. It utilizes channel attention guidance and the CARAFE operator to achieve adaptive fusion of cross-scale semantic features, mitigating feature degradation during upsampling and significantly improving the detection accuracy of the photovoltaic defect detection model for defects of different sizes and shapes. This enhances the scale adaptability and semantic modeling ability of the photovoltaic defect detection model for deformable defects in EL images, further improving its semantic consistency and boundary resolution capabilities under complex texture backgrounds and multi-scale targets.

[0108] Correspondingly, such as Figure 5The diagram illustrates the MEE-YOLO framework. First, PV-EIAE is embedded in the shallow stages of the backbone network to introduce an edge-aware mechanism, improving the response capability of the MEE-YOLO-based photovoltaic defect detection model to weak boundaries and complex textures in EL images. Second, an MCF-FPN structure is adopted in the neck network to enhance the robustness of multi-scale semantic alignment and spatial localization. Finally, for the entire photovoltaic defect detection model, EDRB-MF replaces the existing C3K2 module, achieving improved contextual modeling capabilities while maintaining structural compactness and inference efficiency. Furthermore, within the EDRB-MF structure, a DDR Bottleneck replaces the conventional Bottleneck to strengthen semantic feature expression and structural universality. In other words, this embodiment, based on the improved YOLOv11, constructs a MEE-YOLO photovoltaic defect detection framework integrating PV-EIAE, EDRB-MF, and MCF-FPN to improve the accuracy and robustness of the constructed photovoltaic defect detection model in complex backgrounds and multi-scale defect detection tasks. When using the photovoltaic defect detection model provided in this application to detect defects in photovoltaic cells, it starts from the visual characteristics of EL images and focuses on typical challenges such as blurred defect boundaries, drastic scale changes and severe texture interference, and constructs a detection mechanism that combines structural perception capability and semantic consistency modeling capability.

[0109] In other words, the embodiments of this application structurally optimize YOLOv11 in three aspects: feature extraction, context modeling, and semantic fusion to form a corresponding photovoltaic defect detection model. While ensuring detection accuracy, the real-time performance and robustness of the photovoltaic defect detection model are significantly improved, thus providing a lightweight and efficient solution for photovoltaic defect detection.

[0110] Here, to verify the effectiveness of the photovoltaic defect detection model corresponding to the MEE-YOLO framework proposed in this application embodiment in the defect detection task of photovoltaic cells, PVEL-AD can be used for relevant verification. PVEL-AD contains various real-world photovoltaic cell surface defects and normal samples. Here, because the number of some labeled defects is small or the defect types are not applicable to the defect detection task in this application embodiment, the data for some defect types in PVEL-AD was cleaned and filtered according to the original data description, and finally 3967 images containing 6 typical defects were selected to construct an experimental subset. The 6 typical defects are: cracks, star-shaped cracks, grid lines, black cores, horizontal misalignment, and thick lines. Simultaneously, the experimental subset was divided into a training set and a test set in a 4:1 ratio for subsequent training and performance evaluation of the photovoltaic defect detection model.

[0111] Meanwhile, to comprehensively evaluate the performance of the photovoltaic defect detection model in photovoltaic defect detection tasks, the following commonly used evaluation metrics can be adopted: precision (P), recall (R), mean average precision (mAP)@0.5, and mAP@0.5:0.9; where 0.5 and 0.5:0.9 are the thresholds for intersection-over-union (IoU). Furthermore, to evaluate the efficiency of the photovoltaic defect detection model in deployment, this application also introduces parameters (P) to represent the overall size of the photovoltaic defect detection model, affecting its memory usage and deployment cost; giga floating-point operations per second (GFLOPs) to reflect the computational complexity of the photovoltaic defect detection model (lower values ​​indicate higher efficiency); and frames per second (FPS) to represent the number of image frames processed per second by the photovoltaic defect detection model.

[0112] This embodiment was implemented in a Windows 10 operating system environment, using the Python 3.9.21 programming language and the PyTorch 2.5.1 deep learning framework. The hardware platform configuration included an NVIDIA GeForce RTX 2080Ti (11GB) graphics processing unit (GPU) and an Intel Xeon Gold 5215 central processing unit (CPU) @ 2.50 GHz, accelerated using CUDA 11.8.

[0113] Here, the models involved are trained under a unified configuration, with consistent evaluation metrics. An early stopping strategy is employed during training, terminating training prematurely when mAP@0.5 shows no significant improvement over 20 consecutive training epochs. The optimizer used is Stochastic Gradient Descent (SGD), with a batch size of 16, a total training epochs of 300 (including 5 warm-up epochs), a momentum factor of 0.937, an initial learning rate of 0.013, and a step learning rate decay strategy (Step LR) with a scaling factor of 0.01 and a weight decay coefficient of 0.006. Furthermore, to improve the photovoltaic defect detection model's ability to detect small defects, a mosaic data augmentation mechanism is introduced during the training phase. This mechanism stitches multiple images proportionally to generate new training samples, thereby enhancing the generalization ability and fine-grained target recognition performance of the photovoltaic defect detection model. Meanwhile, in order to comprehensively evaluate the performance of the photovoltaic defect detection model disclosed in the embodiments of this application in photovoltaic cell defect detection, a variety of representative detection algorithms were selected for comparison, and the comparison results are shown in Table 1.

[0114] Table 1 Performance Comparison Results

[0115] ;

[0116] The methods included, as shown in Table 1, are: two-stage detectors, anchor-based single-stage detectors, real-time detection Transformers based on the Transformer architecture, the latest versions of the YOLO series (e.g., Hyper-YOLOn, Hyper-YOLOs, YOLOv12n, YOLOv12s, YOLOv12m), and the best-performing models proposed in recent years (e.g., BAF-Detector, PSA-YOLOv7, C2DEM-YOLO). Table 1 also presents the results for each method in terms of detection accuracy (P, R, mAP), computational complexity (GFLOPs, P), and FPS; arrows indicate performance trends: ↑ indicates higher is better, ↓ indicates lower is better; "-" indicates that this metric is not provided in existing literature.

[0117] Referring again to Table 1, the photovoltaic defect detection model disclosed in this application is compared with traditional detectors: the two-stage detector achieves an mAP@0.5 of 0.891, but suffers from high computational overhead (GFLOPs of 315 and M of 120.7), making it difficult to meet the real-time requirements of industrial scenarios. In contrast, the photovoltaic defect detection model disclosed in this application improves mAP@0.5 by 3.6% while reducing computational complexity by 99.3G, demonstrating a better balance between accuracy and efficiency. The anchor-based single-stage detector relies on a preset anchor frame, which is insufficient for identifying small target defects, and its overall performance is lower than that of the photovoltaic defect detection model proposed in this paper.

[0118] Meanwhile, a comparison between the photovoltaic defect detection model disclosed in this application and the YOLO series models is as follows:

[0119] The photovoltaic defect detection model of this application achieves an mAP@0.5 of 0.927 while maintaining a low parameter count (M = 6.9), which is significantly better than YOLOv12n (mAP@0.5 = 0.883) and YOLOv12s (mAP@0.5 = 0.892). At the same time, the inference speed FPS is 169.5, which is better than the YOLO series models such as Hyper-YOLOn, Hyper-YOLOs, YOLOv12n, YOLOv12s, and YOLOv12m, demonstrating good performance in both speed and accuracy.

[0120] Defect detection using the photovoltaic defect detection model disclosed in the embodiments of this application is compared with defect detection using recent advanced methods:

[0121] Compared to BAF-Detector (mAP@0.5=0.806), PSA-YOLOv7 (mAP@0.5=0.829), and C2DEM-YOLO (mAP@0.5=0.923), the photovoltaic defect detection model in this application achieves comprehensive superiority in detection accuracy (mAP@0.5=0.927). Furthermore, while maintaining relatively low model complexity, its inference speed is further improved to 169.5, exceeding the current best-performing model, C2DEM-YOLO (FPS=157.6), demonstrating superior real-time performance and generalization ability.

[0122] In summary, the photovoltaic defect detection model built on the MEE-YOLO framework in this application outperforms current mainstream detection methods in terms of detection accuracy, inference speed, and computational efficiency. By integrating the edge information assisted extraction module (PV-EIAE), the context semantic modeling module (EDRB-MF), and the multi-scale fusion network (MCF-FPN), the feature modeling capability of the photovoltaic defect detection model in complex texture backgrounds and multi-scale defect scenarios can be significantly improved. This photovoltaic defect detection model combines high accuracy and high efficiency, and has good deployability and industrial adaptability, making it particularly suitable for photovoltaic module defect detection tasks that have dual requirements for real-time performance and detection accuracy.

[0123] Furthermore, to verify the independent contribution and synergistic gain of each module (PV-EIAE, EDRB-MF, and MCF-FPN) to the photovoltaic defect detection performance proposed in this application, eight ablation experiments were designed and implemented on PVEL-AD. Table 2 lists the performance under different module combinations, i.e., the module performance comparison of the improved YOLOv11 on PVEL-AD.

[0124] Table 2 Ablation Experiment Results

[0125] ;

[0126] Table 2 evaluates the impact of different module combinations on detection performance, with metrics including: precision (P), recall (R), mAP@0.5, mAP@0.5:0.9, computational complexity (GFLOPs), number of parameters (P), and inference speed (FPS). Best performance is indicated in bold, and second-best results are indicated by underline. "" indicates that the module is included in the model, and "-" indicates that it is not used. Arrows indicate performance trends: ↑ indicates higher values ​​are better, and ↓ indicates lower values ​​are better. Table 2 shows the impact on individual modules: using YOLOv11n as the baseline model (P=0.841, R=0.808, mAP@0.5=0.888, mAP@0.5:0.9=0.559), comparative experiments were conducted by introducing the PV-EIAE, EDRB-MF, and MCF-FPN modules respectively. After introducing PV-EIAE, P increased by 2. The improvement of R by 0% and mAP@0.5 by 1.1% verifies the effectiveness of the edge perception mechanism in the photovoltaic defect detection model in terms of fine-grained defect perception and robustness in complex backgrounds. The addition of EDRB-MF improves R by 5.3%, but P decreases slightly (-0.6%), indicating that while the multi-scale expansion structure enhances the ability to model remote contexts, it may also introduce some background interference. After introducing MCF-FPN, P improves by 1.5% and mAP@0.5:0.9 improves by 1.3%, demonstrating its advantages in semantic recombination and multi-scale fusion.

[0127] Meanwhile, regarding the synergistic effect of the combined modules: further evaluation of the synergistic gains brought by the combination of each module: the PV-EIAE+EDRB-MF combination improved R by 3.4% and mAP@0.5 by 1.5%, demonstrating the complementary effect of edge enhancement and semantic modeling; the EDRB-MF+MCF-FPN combination optimized context modeling while maintaining multi-scale semantic perception capabilities, improving R by 5.1%; the PV-EIAE+MCF-FPN combination performed better in terms of P (0.864) and mAP@0.5 (0.895), indicating that the synergistic enhancement effect in terms of detail preservation and scale fusion is significant.

[0128] In other words, the complete model constructed by integrating the three modules PV-EIAE, EDRB-MF, and MCF-FPN—the photovoltaic defect detection model—achieves optimal detection performance under comprehensive performance indicators of P=0.865, R=0.853, mAP@0.5=0.927, and mAP@0.5:0.9=0.581. While ensuring improved detection accuracy, the complete photovoltaic defect detection model maintains a high inference speed (169.5 FPS) and a computational complexity of only 2.345 GFLOPs, far lower than traditional detectors such as Faster R-CNN (315.0 GFLOPs), demonstrating excellent computational efficiency.

[0129] In summary, PV-EIAE effectively improves the robustness of the photovoltaic defect detection model in edge perception and complex backgrounds, EDRB-MF enhances the model's multi-scale context modeling and semantic representation capabilities, and MCF-FPN optimizes its cross-scale feature fusion and detail preservation performance. The synergistic integration of these three technologies significantly enhances the photovoltaic defect detection model's performance in complex scenarios, thus validating the practicality and engineering deployment value of the photovoltaic defect detection model disclosed in this application.

[0130] Here you can refer to Figure 6 The image shows a comparison of the detection results of the photovoltaic defect detection model disclosed in this application and the YOLOv11n model under different defect types; the red boxes mark the defect areas that the YOLOv11n model failed to recognize but that the photovoltaic defect detection model provided in this application successfully detected. Figure 6 The first row of the image is the original image, the second row is the detection result output by YOLOv11n, and the third row is the detection result output by the photovoltaic defect detection model in this application. Figure 6Columns 1 and 3 show that YOLOv11n exhibits significant missed detections when identifying small-scale defects such as microcracks and localized grid breaks. In contrast, the photovoltaic defect detection model in this application can more accurately locate fine-grained structures, thanks to the multi-scale edge perception mechanism introduced by PV-EIAE, which effectively enhances the ability to represent edge features in complex backgrounds. Figure 6 In the second column of samples, YOLOv11n exhibited significant false positives and bounding box overlap, reflecting its limited ability to discriminate targets in complex defect structures. In contrast, the photovoltaic defect detection model in this application, leveraging the remote context modeling capability constructed by EDRB-MF, significantly improves semantic discrimination and region awareness performance. Furthermore, in Figure 6 In the fourth and fifth columns of the samples, YOLOv11n struggled to accurately distinguish between lattice textures and actual defect regions, resulting in missed detections and location errors. In contrast, the photovoltaic defect detection model in this application utilizes MCF-FPN to achieve multi-level semantic information fusion and detailed feature reconstruction, effectively mitigating the detection difficulties caused by weak responses to small targets and texture interference. In summary, Figure 6 The visualization comparison results fully demonstrate that the photovoltaic defect detection model in this application is significantly better than YOLOv11n in terms of fine-grained defect identification, complex background adaptation and cross-scale target detection, and verify the effectiveness and robustness of the synergistic enhancement of each module.

[0131] In other words, this application addresses the limitations in feature extraction and low semantic fusion efficiency caused by lattice texture interference and multi-scale variation of defects in EL images. It constructs a MEE-YOLO photovoltaic defect detection framework based on an improved YOLOv11, integrating edge enhancement, context modeling, and multi-scale feature fusion strategies to significantly improve detection accuracy and robustness. The designed photovoltaic defect detection model incorporates PV-EIAE, which enhances its ability to perceive weak boundaries and fine-grained defects; EDRB-MF, which effectively expands the receptive field of the photovoltaic defect detection model while maintaining its computational efficiency; and MCF-FPN, which improves the fusion efficiency of information at different scales. These three components work together to construct a defect detection model that combines high efficiency and accuracy—the photovoltaic defect detection model.

[0132] Furthermore, experiments on PVEL-AD demonstrate that the photovoltaic defect detection model provided in this application achieves 92.7% accuracy and 58.1% accuracy on mAP@0.5 and mAP@0.5:0.95, respectively, with an inference speed of 169.5 and a computational complexity of only 2.345 GFLOPs, outperforming current mainstream detection methods. In other words, the experimental results show that the photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion provided in this application achieves a good balance between accuracy and efficiency, exhibiting strong engineering applicability and the potential for practical deployment in industrial photovoltaic inspection systems.

[0133] Overall, the photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion provided in this application effectively improves edge perception, context modeling, and multi-scale feature fusion, thereby enhancing the robustness of detection in complex EL imaging environments and providing a practically valuable solution for efficient and reliable defect detection in photovoltaic systems.

[0134] It should be noted that, in the embodiments of this application, if the above-mentioned photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a magnetic disk, or an optical disk. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0135] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0136] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several methods provided in this application can be arbitrarily combined to obtain new method embodiments without conflict.

[0137] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A photovoltaic cell defect detection method based on edge enhancement and multi-scale feature fusion, characterized in that, The method comprises: acquiring an electroluminescence (EL) image of a photovoltaic cell; extracting multi-scale context features from the EL image in parallel through parallel convolution branches with different dilation rates in an enhanced dilated convolution residual block of a backbone network in a photovoltaic defect detection model to obtain multiple groups of initial features; wherein the backbone network in the photovoltaic defect detection model is integrated with a photovoltaic edge information auxiliary extractor and the enhanced dilated convolution residual block; the photovoltaic edge information auxiliary extractor introduces a hierarchical edge perception mechanism; integrating the multiple groups of initial features through a semantic residual fusion module in the enhanced dilated convolution residual block to obtain integrated features, and performing residual connection on the integrated features and the EL image to obtain an original visual feature image with high-level semantic information; extracting multi-scale edge features from the original visual feature image through a multi-scale edge information generator in the photovoltaic edge information auxiliary extractor to obtain a multi-scale edge feature image set; gradually fusing the multi-scale edge feature image set and the original visual feature image through multiple cascaded edge feature transmission fusion modules in the photovoltaic edge information auxiliary extractor to obtain multi-scale feature images; wherein the number of modules corresponding to the multiple cascaded edge feature transmission fusion modules is the same as the number of images included in the multi-scale edge feature image set; based on each scale feature image, determining global average channel-level features and global maximum channel-level features of the scale feature image through global average pooling layers and global maximum pooling layers in a feature selection module in a multi-level context feature fusion pyramid network; wherein the multi-level context feature fusion pyramid network is a neck network of the photovoltaic defect detection model; fusing the global average channel-level features and the global maximum channel-level features through fully connected layers and nonlinear activation functions in the feature selection module to obtain a channel attention image of the scale feature image; weighting the scale feature image based on the channel attention image of the scale feature image through an adaptive attention module in the feature selection module to obtain an enhanced feature image corresponding to each scale feature image; determining a multi-scale enhanced feature image composed of the enhanced feature images corresponding to each scale feature image; performing multi-level semantic feature fusion on each scale enhanced feature image in the multi-scale enhanced feature image through an adaptive feature fusion module in the multi-level context feature fusion pyramid network to obtain a multi-scale semantic feature enhanced image; performing defect detection on the multi-scale semantic feature enhanced image through a head network of the photovoltaic defect detection model to obtain a defect type and a defect position of the photovoltaic cell.

2. The method of claim 1, wherein, The method comprises: The Sobel operator in the multi-scale edge information generator is used to determine a horizontal gradient component in a horizontal direction and a vertical gradient component in a vertical direction of the original visual feature image, and to construct an initial edge feature image according to the horizontal gradient component and the vertical gradient component; The maximum pooling layer in the multi-scale edge information generator is used to construct a pyramid edge representation feature image corresponding to the initial edge feature image; The convolution layer in the multi-scale edge information generator is used to perform channel compression and feature alignment on the pyramid edge representation feature image to obtain the multi-scale edge feature image set.

3. The method of claim 1, wherein, The multi-scale feature image is obtained by performing step-by-step fusion on the multi-scale edge feature image set and the original visual feature image through the plurality of cascaded edge feature transmission fusion modules in the photovoltaic edge information auxiliary extractor, including: For each edge feature transmission fusion module, the edge feature transmission module in the edge feature transmission fusion module is used to select a scale edge feature image corresponding to a cascade to which the edge feature transmission fusion module belongs from the multi-scale edge feature image set; The enhanced dilated convolution residual block in the edge feature transmission fusion module is used to perform feature extraction on the intermediate feature image to obtain an efficient feature representation image of the edge feature transmission fusion module; The edge feature transmission fusion module is the first edge feature transmission fusion module in the plurality of cascaded edge feature transmission fusion modules, and the original visual feature image is the intermediate feature image; when the edge feature transmission fusion module is not the first edge feature transmission fusion module in the plurality of cascaded edge feature transmission fusion modules, the intermediate feature image is an output feature image corresponding to the edge feature transmission fusion module of the previous stage; The convolution edge fusion module in the edge feature transmission fusion module is used to perform feature fusion on the scale edge feature image and the efficient feature representation image to obtain an output feature image corresponding to the edge feature transmission fusion module; The multi-scale feature image is determined by the output feature images corresponding to the edge feature transmission fusion modules of each stage.

4. The method of claim 3, wherein, The convolution edge fusion module in the edge feature transmission fusion module is used to perform feature fusion on the scale edge feature image and the efficient feature representation image to obtain an output feature image corresponding to the edge feature transmission fusion module, including: The concatenation layer in the convolution edge fusion module is used to perform feature concatenation on the scale edge feature image and the efficient feature representation image to obtain a concatenated feature image; The convolution layer in the convolution edge fusion module is used to sequentially perform channel compression, local perception enhancement operation, and fusion reconstruction on the concatenated feature image to obtain the output feature image corresponding to the edge feature transmission fusion module.

5. The method of claim 1, wherein, The construction process of the photovoltaic defect detection model includes: An initial detection model constructed by YOLOv11 is obtained; Integrate the photovoltaic edge information auxiliary extractor in the backbone network of the initial detection model, and replace the C3K2 module for feature extraction in the initial detection model with the enhanced dilated convolution residual block in a synchronous manner to obtain an intermediate detection model; Change the neck network in the intermediate detection model into the multi-level context feature fusion pyramid network to obtain an improved initial detection model; Take a photovoltaic cell defect anomaly detection dataset as a training set, train the improved initial detection model by minimizing a loss function, and obtain the photovoltaic defect detection model.

Images