A high-performance solar cell appearance defect detection method
By introducing a bidirectional path feature pyramid and a group attention module into the YOLOv5 model, the problems of complex backgrounds, small and weak defects, and varying scales in solar cell appearance defect detection are solved, achieving efficient defect detection, improving performance and reducing model complexity.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-05
- Publication Date
- 2026-04-14
AI Technical Summary
Existing convolutional neural network (CNN) detection methods face challenges in detecting defects in solar cells, such as complex backgrounds, tiny and weak defect targets, and varying defect scales, resulting in insufficient detection capabilities.
A bidirectional path feature pyramid structure and a group attention module are embedded into the YOLOv5 model. Through top-down forward feature fusion and bottom-up reverse feature fusion, combined with CBS, C3, SPPF and GAM modules, multi-scale feature extraction and feature refinement are achieved, redundant background information is suppressed and contextual information is preserved.
It significantly improves the detection capability for tiny and weak defects, with a 3% performance improvement, fewer parameters, faster inference speed, and improved defect detection performance in complex backgrounds.
Smart Images
Figure CN116863299B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of solar cell defect detection technology, specifically a method for detecting appearance defects in high-performance solar cells. Background Technology
[0002] During the production of solar cells, factors such as screen printing, etching, and mechanical stress inevitably lead to appearance defects, including edge chipping, grid breakage, paste leakage, and dirty cells. These defects significantly affect the photoelectric conversion efficiency of solar cells, and consequently, the power generation efficiency of the photovoltaic system. Therefore, appearance defect detection is a crucial step in the solar cell manufacturing process.
[0003] Detection methods based on convolutional neural networks (CNNs) have the advantages of high recognition accuracy, fast speed and strong robustness, and have been widely used in the field of defect detection. However, there are still huge challenges in using CNNs to detect appearance defects of solar cells, specifically: (1) Complex background: The grid lines of polycrystalline silicon solar cells are dense and the lattice shape is irregular, resulting in low distinction between complex texture background and defect area; in addition, the color intensity of polycrystalline silicon solar cells is obviously uneven, and the uneven change of background and defect color intensity is one of the main factors affecting defect detection performance; (2) Small and weak defect target: The pixel area of the defect area is small, and the area of the defect accounts for only 0.0015% of the entire image, and the maximum does not exceed 1.56%; (3) Different defect scales: There are many types of appearance defects of solar cells, and the scales of different types of defects are different; in addition, there are also significant scale differences in defects of the same type. For example, the pixel area range of the defect "dirty sheet" is 440 to 75624, and the maximum pixel area of the defect "leaking paste" is 4350, and the minimum is only 192.
[0004] For problems such as complex backgrounds, tiny and weak defect targets, and defects of varying scales, existing CNN detection methods are clearly insufficient. Therefore, this invention proposes a high-performance method for detecting appearance defects in solar cells, enabling efficient detection of these defects. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a method for detecting appearance defects in high-performance solar cells.
[0006] The present invention solves the aforementioned technical problem by adopting the following technical solution:
[0007] A method for detecting appearance defects in high-performance solar cells, characterized by comprising the following steps:
[0008] Step 1: Obtain images of defects in the appearance of the solar cell;
[0009] Step 2: Constructing an appearance defect detection model; the appearance defect detection model consists of three parts: feature extraction, feature fusion, and classification and regression; the input image is first processed by the feature extraction part to extract features, obtaining multi-scale features; the multi-scale features are then fused in the feature fusion part, which uses a bidirectional path feature pyramid, including a top-down forward feature fusion path and a bottom-up reverse feature fusion path, strengthening the fusion features of the same scale in the two paths to obtain multiple strengthened fusion features; each strengthened fusion feature is then classified and regressed in the corresponding detection head of the classification and regression part to obtain prediction results at different scales;
[0010] Step 3: Input the images of solar cell appearance defects into the initialized appearance defect detection model for training, and then deploy the trained appearance defect detection model into the detection equipment for the detection of solar cell appearance defects.
[0011] Furthermore, in the forward feature fusion path, the output features of the second C3 module in the feature extraction part are passed through a CBS module to obtain the first fused feature of the forward feature fusion path; the first fused feature of the forward feature fusion path is downsampled and then concatenated with the output features of the third C3 module in the feature extraction part; the concatenated feature is then passed through a C3 module to obtain the second fused feature of the forward feature fusion path; the second fused feature of the forward feature fusion path is then passed through a CBS module and a downsampling operation in sequence, and then concatenated with the output features of the SPPF module in the feature extraction part; the concatenated feature is then passed through a C3 module to obtain the third fused feature of the forward feature fusion path.
[0012] In the reverse feature fusion path, the output features of the SPPF module in the feature extraction part are passed through a CBS module to obtain the first fused feature of the reverse feature fusion path. After an upsampling operation, the first fused feature of the reverse feature fusion path is concatenated with the output features of the third C3 module in the feature extraction part. The concatenated feature is then passed through a C3 module to obtain the second fused feature of the reverse feature fusion path. The second fused feature of the reverse feature fusion path is then passed through a CBS module and an upsampling operation in sequence, and then concatenated with the output features of the second C3 module in the feature extraction part. The concatenated feature is then passed through a C3 module to obtain the third fused feature of the reverse feature fusion path.
[0013] The third fusion feature of the reverse feature fusion path is fused with the first fusion feature of the forward feature fusion path after passing through a GAM module to obtain the first enhanced fusion feature. The second fusion feature of the forward feature fusion path and the second fusion feature of the reverse feature fusion path are fused with each of the GAM modules to obtain the second enhanced fusion feature. The third fusion feature of the forward feature fusion path is fused with the first fusion feature of the reverse feature fusion path after passing through a GAM module to obtain the third enhanced fusion feature. These three enhanced fusion features are the output features of the feature fusion part.
[0014] The C3 module mentioned above uses the CBS module and BottleNeck module as basic units for feature fusion; the CBS module includes convolutional layers, normalization layers and activation layers, and the GAM module is a group attention module.
[0015] Furthermore, in the CBS module, the input features of the CBS module pass through convolutional layers and normalization layers in sequence to obtain normalized features; the normalized features pass through activation layers and are then multiplied by themselves to obtain the output features of the CBS module; in the C3 module, the input features of the C3 module pass through the CBS module and BottleNeck module in sequence, and are then concatenated with the features obtained by passing the input features through a CBS module. The concatenated features then pass through a CBS module to obtain the output features of the C3 module.
[0016] Furthermore, in the GAM module, the input features are split into multiple groups of features according to the number of channels through a reconstruction operation. Each group of features is then passed through an attention branch and concatenated. The concatenated features are then added to the features obtained from the input features of the GAM module after passing through an attention branch to obtain the output features of the GAM module. Specifically, the input features of the attention branch are subjected to average pooling and max pooling operations. The features obtained from the two pooling operations are concatenated and then passed through a convolutional layer and an activation layer in sequence. Finally, they are multiplied with the input features of the attention branch to obtain the output features of the attention branch.
[0017] Furthermore, the feature extraction part includes five CBS modules, four C3 modules, and one SPPF module; the first two CBS modules are connected together, and each of the second to fifth CBS modules is connected to a C3 module, with the SPPF module located after the fourth C3 module; the output features of the second C3 module, the third C3 module, and the SPPF module are the multi-scale features extracted by the feature extraction part.
[0018] Compared with the prior art, the beneficial effects of the present invention are:
[0019] 1. This invention embeds the proposed bidirectional path feature pyramid structure and group attention module into the YOLOv5 model to obtain an efficient appearance defect detection model. This significantly improves the model's ability to detect small and weak defects in complex backgrounds and demonstrates superior detection performance. Compared with the original YOLOv5 model, the appearance defect detection model of this invention improves performance by 3%, has fewer parameters, and faster inference speed.
[0020] 2. This invention addresses the issues of minute and weak defects and defects of varying scales by proposing a bidirectional path feature pyramid structure in the feature fusion section. A bottom-up reverse feature fusion path achieves the fusion of deep features to shallow features, while a top-down forward feature fusion path achieves the fusion of shallow features to deep features. Features of the same scale from both fusion paths are then enhanced through fusion, giving the model a strong multi-scale feature extraction capability. BP-FPN changes the traditional serial feature fusion method of the feature pyramid to parallel fusion, improving the problem of feature information duplication and redundant superposition during feature fusion. By fusing features of the same scale, it preserves the characteristics of minute and weak defects in shallow features, significantly improving the model's ability to detect minute and weak targets.
[0021] 3. To address the problem of accurate defect location detection in complex backgrounds, this invention proposes a group attention module to refine features, suppress redundant background information, retain more contextual information, highlight the defect location, and improve the detection capability of small and weak defects in complex backgrounds. Attached Figure Description
[0022] Figure 1 This is a structural diagram of the appearance defect detection model of the present invention;
[0023] Figure 2 Here is a structural diagram of the CBS module;
[0024] Figure 3 Here is the structure diagram of module C3;
[0025] Figure 4 This is a structural diagram of an SPPF module;
[0026] Figure 5 This is a structural diagram of the GAM module. Detailed Implementation
[0027] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments, but this does not limit the scope of protection of this application.
[0028] This invention relates to a method for detecting appearance defects in high-performance solar cells (hereinafter referred to as the method, see [link]). Figures 1 to 5 The specific steps include the following:
[0029] Step 1: Establish a dataset of solar cell appearance defects;
[0030] To obtain solar cell appearance images using industrial cameras, it is necessary to include defect categories required for quality inspection and label the defect areas in the images with their locations and categories to obtain a solar cell appearance defect dataset. The solar cell appearance defect dataset is then randomly divided into a training set and a test set in an 8:2 ratio.
[0031] Step 2: Construct an appearance defect detection model;
[0032] See Figure 1 The appearance defect detection model consists of three parts: feature extraction, feature fusion, and classification and regression. When the input image enters the appearance defect detection model, it first performs feature extraction in the feature extraction part to obtain multi-scale features. The multi-scale features are then fused in the feature fusion part to obtain multiple enhanced fused features. Each enhanced fused feature is then classified and regressed in the corresponding detection head of the classification and regression part to obtain prediction results at different scales.
[0033] The feature extraction part includes five CBS modules, four C3 modules, and one SPPF module; the first two CBS modules are connected together, and each of the second to fifth CBS modules is connected to a C3 module. The SPPF module is located after the fourth C3 module. The output features of the second C3 module, the third C3 module, and the SPPF module are the multi-scale features extracted by the feature extraction part.
[0034] The feature fusion part adopts a bidirectional path feature pyramid structure (BP-FPN), which includes a top-down forward feature fusion path and a bottom-up reverse feature fusion path. The forward feature fusion path realizes the fusion from shallow features to deep features, and the reverse feature fusion path realizes the fusion from deep features to shallow features. The fused features of the same scale in the two paths are enhanced and fused to obtain a total of three enhanced fused features, which are the output features of the feature fusion part.
[0035] In the forward feature fusion path, the output features of the second C3 module in the feature extraction part are passed through a CBS module to obtain the first fused feature of the forward feature fusion path. After a downsampling operation, the first fused feature of the forward feature fusion path is concatenated with the output features of the third C3 module in the feature extraction part. The concatenated feature is then passed through another C3 module to obtain the second fused feature of the forward feature fusion path. The second fused feature of the forward feature fusion path is then passed through a CBS module and a downsampling operation in sequence, and then concatenated with the output features of the SPPF module in the feature extraction part. The concatenated feature is then passed through another C3 module to obtain the third fused feature of the forward feature fusion path.
[0036] In the reverse feature fusion path, the output features of the SPPF module in the feature extraction part are passed through a CBS module to obtain the first fused feature of the reverse feature fusion path. After an upsampling operation, the first fused feature of the reverse feature fusion path is concatenated with the output features of the third C3 module in the feature extraction part. The concatenated feature is then passed through a C3 module to obtain the second fused feature of the reverse feature fusion path. The second fused feature of the reverse feature fusion path is then passed through a CBS module and an upsampling operation in sequence, and then concatenated with the output features of the second C3 module in the feature extraction part. The concatenated feature is then passed through a C3 module to obtain the third fused feature of the reverse feature fusion path.
[0037] Finally, the fusion features of the same scale in the forward and reverse feature fusion paths are enhanced through fusion. Specifically, the third fusion feature of the reverse feature fusion path is fused with the first fusion feature of the forward feature fusion path after passing through a GAM module, resulting in the first enhanced fusion feature. Similarly, the second fusion feature of the forward and reverse feature fusion paths are fused with the first fusion feature of the reverse feature fusion path after passing through a GAM module, resulting in the second enhanced fusion feature. Finally, the third fusion feature of the forward feature fusion path is fused with the first fusion feature of the reverse feature fusion path after passing through a GAM module, resulting in the third enhanced fusion feature. BP-FPN changes the traditional serial feature fusion method of the feature pyramid to parallel fusion, improving the problem of feature duplication and redundant superposition in feature fusion, highlighting the fusion of features of the same scale, obtaining richer spatial information, and improving the model's ability to detect small and weak targets.
[0038] like Figure 2 As shown, the CBS module includes a convolutional layer, a normalization layer, and an activation layer. The input features of the CBS module pass through the convolutional layer and the normalization layer in sequence to obtain normalized features. After the normalized features pass through the activation layer, they are multiplied by themselves to obtain the output features of the CBS module.
[0039] likeFigure 3 As shown, the C3 module is used for feature fusion and includes a CBS module and a BottleNeck module. The input features of the C3 module pass through the CBS module and the BottleNeck module sequentially, and are then concatenated with the features obtained from passing through a CBS module. The concatenated features then pass through another CBS module to obtain the output features of the C3 module. The main function of the C3 module is to increase the depth of the network and improve the feature extraction capability. The BottleNeck module includes three sequential convolutional layers. The first two convolutional layers are each followed by an activation layer. The output features of the third convolutional layer are concatenated with the input features of the BottleNeck module. The concatenated features then pass through an activation layer to obtain the output features of the BottleNeck module.
[0040] like Figure 4 As shown, the SPPF module includes a CBS module and a max pooling layer. The input features of the SPPF module pass through a CBS module, then through three max pooling layers in sequence. The features obtained from the three max pooling layers are then concatenated with the output features of the CBS module. The concatenated features then pass through another CBS module to obtain the output features of the SPPF module. The main function of the SPPF module is to solve the problem of different input image sizes.
[0041] In the bidirectional path feature pyramid structure, the group attention (GAM) module refines the features, suppresses redundant background information, and retains more contextual information, thereby improving the detection capability of small and weak target defects in complex backgrounds; for example... Figure 5 As shown, the input features of the GAM module are split into multiple groups according to the number of channels through a reconstruction operation. Each group of features is then processed through an attention branch and concatenated. The concatenated features are then added to the features obtained from the input features of the GAM module after processing through an attention branch, resulting in the output features of the GAM module. Specifically, the input features of the attention branch undergo average pooling and max pooling operations. The features obtained from these two pooling operations are concatenated, then sequentially processed through a convolutional layer and an activation layer, and finally multiplied with the input features of that attention branch to obtain its output features. Since shallow features contain redundant background information, the GAM module uses an attention mechanism to suppress this background information. Simultaneously, to prevent the target region from being suppressed as background, the GAM module groups the features and refines them to retain more contextual information, thereby improving detection performance.
[0042] Step 3: Training and testing of the appearance defect detection model
[0043] (1) Sample data augmentation
[0044] Mosic data augmentation is used on the training set divided in the first step. One input image of the appearance of a solar cell is selected, and three more images are randomly selected. These four images are cropped and stitched together to obtain a new image, which is the training sample.
[0045] (2) Parameter settings
[0046] Initialize model parameters, including but not limited to: network weights and biases; set the maximum number of iterations (epochs) to 200; set the learning rate to 0.001 for the first 50 epochs and 0.0001 for the next 150 epochs; set the learning rate decrease factor to 0.1 and the weight decay of the regularization term to 0.0005.
[0047] (3) Training and Testing
[0048] The training samples are input into the initialized appearance defect detection model. The loss is calculated using the loss function, and the model is optimized through backpropagation until the loss converges. The loss function includes localization loss, confidence loss, and classification loss. The classification loss and confidence loss use binary cross-entropy loss (BCE Loss), and the localization loss uses CIoU loss.
[0049] The test set is input into the trained appearance defect detection model for testing to verify the model's defect detection capability. The model's detection results are compared with the labels to evaluate the model's performance in detecting appearance defects.
[0050] Step 4: Deploy the trained appearance defect detection model onto the detection equipment to detect appearance defects in solar cells.
[0051] To verify the effectiveness of the method of the present invention, the method of the present invention was compared with common defect detection models, and the detection results are shown in Table 1.
[0052] Table 1 Statistical analysis of detection results for different models
[0053]
[0054] As can be seen from the table, the average detection accuracy of the method of the present invention for the four types of defects is higher than that of the other common defect detection models, which verifies the effectiveness of the method of the present invention.
[0055] Any aspects not covered in this invention are applicable to existing technologies.
Claims
1. A method for detecting appearance defects in high-performance solar cells, characterized in that, The method includes the following steps: Step 1: Obtain images of defects in the appearance of the solar cell; Step 2: Constructing an appearance defect detection model; the appearance defect detection model consists of three parts: feature extraction, feature fusion, and classification and regression; the input image is first processed by the feature extraction part to extract features, obtaining multi-scale features; the multi-scale features are then fused in the feature fusion part, which uses a bidirectional path feature pyramid, including a top-down forward feature fusion path and a bottom-up reverse feature fusion path, strengthening the fusion features of the same scale in the two paths to obtain multiple strengthened fusion features; each strengthened fusion feature is then classified and regressed in the corresponding detection head of the classification and regression part to obtain prediction results at different scales; In the forward feature fusion path, the output features of the second C3 module in the feature extraction part are passed through a CBS module to obtain the first fused feature of the forward feature fusion path. After a downsampling operation, the first fused feature of the forward feature fusion path is concatenated with the output features of the third C3 module in the feature extraction part. The concatenated feature is then passed through another C3 module to obtain the second fused feature of the forward feature fusion path. The second fused feature of the forward feature fusion path is then passed through a CBS module and a downsampling operation in sequence, and then concatenated with the output features of the SPPF module in the feature extraction part. The concatenated feature is then passed through another C3 module to obtain the third fused feature of the forward feature fusion path. In the reverse feature fusion path, the output features of the SPPF module in the feature extraction part are passed through a CBS module to obtain the first fused feature of the reverse feature fusion path. After an upsampling operation, the first fused feature of the reverse feature fusion path is concatenated with the output features of the third C3 module in the feature extraction part. The concatenated feature is then passed through a C3 module to obtain the second fused feature of the reverse feature fusion path. The second fused feature of the reverse feature fusion path is then passed through a CBS module and an upsampling operation in sequence, and then concatenated with the output features of the second C3 module in the feature extraction part. The concatenated feature is then passed through a C3 module to obtain the third fused feature of the reverse feature fusion path. The third fusion feature of the reverse feature fusion path is fused with the first fusion feature of the forward feature fusion path after passing through a GAM module to obtain the first enhanced fusion feature. The second fusion feature of the forward feature fusion path and the second fusion feature of the reverse feature fusion path are fused with each of the GAM modules to obtain the second enhanced fusion feature. The third fusion feature of the forward feature fusion path is fused with the first fusion feature of the reverse feature fusion path after passing through a GAM module to obtain the third enhanced fusion feature. These three enhanced fusion features are the output features of the feature fusion part. The C3 module mentioned above uses the CBS module and BottleNeck module as basic units for feature fusion; the CBS module includes convolutional layers, normalization layers and activation layers, and the GAM module is a group attention module; Step 3: Input the images of solar cell appearance defects into the initialized appearance defect detection model for training, and then deploy the trained appearance defect detection model into the detection equipment for the detection of solar cell appearance defects.
2. The method for detecting appearance defects in high-performance solar cells according to claim 1, characterized in that, In the CBS module, the input features of the CBS module pass through convolutional layers and normalization layers in sequence to obtain normalized features; The normalized features are passed through the activation layer and then multiplied by themselves to obtain the output features of the CBS module; In module C3, the input features of module C3 pass through modules CBS and BottleNeck in sequence, and are then concatenated with the features obtained by passing the input features through a CBS module. The concatenated features then pass through a CBS module to obtain the output features of module C3.
3. The method for detecting appearance defects in high-performance solar cells according to claim 1 or 2, characterized in that, In the GAM module, the input features are split into multiple groups according to the number of channels through a reconstruction operation. Each group of features is then passed through an attention branch and concatenated. The concatenated features are then added to the features obtained from the input features of the GAM module after passing through an attention branch to obtain the output features of the GAM module. Specifically, the input features of the attention branch are subjected to average pooling and max pooling operations. The features obtained from the two pooling operations are concatenated and then passed through a convolutional layer and an activation layer in sequence. Finally, they are multiplied with the input features of the attention branch to obtain the output features of the attention branch.
4. The method for detecting appearance defects in high-performance solar cells according to claim 1, characterized in that, The feature extraction part includes five CBS modules, four C3 modules, and one SPPF module; the first two CBS modules are connected together, and each of the second to fifth CBS modules is connected to a C3 module. The SPPF module is located after the fourth C3 module. The output features of the second C3 module, the third C3 module, and the SPPF module are the multi-scale features extracted by the feature extraction part.