A solar cell appearance defect detection method

Abstract

The application discloses a solar cell appearance defect detection method, the model used by the method comprises a multi-scale feature extraction network, a cross-stage parallel feature fusion network and a classification regression network; the multi-scale feature extraction network is formed by connecting a plurality of multi-scale residual feature extraction modules; the feature maps output by the third, second and first multi-scale residual feature extraction modules are multi-scale feature maps obtained by the multi-scale feature extraction network; the cross-stage parallel feature fusion network is divided into two stages in the horizontal direction and three parallel branches in the vertical direction; information of each branch is fused with each other, so that the expression ability of the features is stronger. The multi-scale feature extraction network can retain defect features of different scales and avoid loss of tiny and weak defect information caused by pooling, convolution and other operations along with network extension; the cross-stage parallel feature fusion network combines bidirectional multi-scale features between shallow layers and deep layers, so that loss of tiny and weak defect information in the shallow layers in the feature fusion process is avoided, and therefore, the detection capability of the model for tiny and weak defects in the photovoltaic cell appearance is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of solar cell defect detection, and particularly relates to a solar cell appearance defect detection method. BACKGROUND

[0002] Due to process factors, human operation errors and other reasons, defects such as edge collapse, broken grid, slurry leakage and dirty pieces inevitably occur in the production process of solar cell pieces, which greatly affect the photoelectric efficiency of photovoltaic cell pieces, and further affect the stability of photovoltaic cell modules and even the entire photovoltaic system, so defect detection plays a crucial role in the production process of solar cell pieces.

[0003] Traditional manual detection is easily affected by external environment or human factors, has poor anti-interference performance, and has high detection cost. The detection method based on machine vision includes traditional image processing algorithm and target detection method based on deep learning. The pre-processing part of traditional image processing algorithm such as linear regression, decision tree algorithm and support vector machine is relatively complex, and the recognition effect of defects is poor, so it does not have industrial application value. The target detection method based on deep learning is widely used due to its fast recognition speed, high precision and strong robustness, and common detection models include YOLOv5, YOLOv7, FasterRCNN, etc.

[0004] The surface defects of solar cell pieces have obvious "weakness" and "smallness" characteristics, mainly in the following three aspects: 1) uneven background interference: the color intensity of the cell piece is obviously uneven, and the multi-crystalline silicon solar cell piece has a complex texture background. 2) less defect information: the pixel area of the defect accounts for a very small proportion of the overall image, the smallest is only 0.0015%, and the largest is only 1.56%. 3) large defect size variation: the size difference of different defect types is large, and even the same type of defect may have significant size difference. Therefore, a large number of defects in the solar cell piece defects are small target defects. For the target detection task of deep learning, small target defect detection has always been a heavy and difficult point. On the one hand, with the deepening of convolutional neural network, feature loss problem will occur through downsampling, pooling and other operations, especially the loss of small and weak feature information. On the other hand, small target defects have small size and weak contrast, and contain less information, so the small target defect feature expression ability is not strong, and it is easy to be ignored in the learning process, which leads to that the detection accuracy of the model for small scale targets is usually only half of that for large scale targets. SUMMARY

[0005] In view of the deficiencies of the prior art, the technical problem to be solved by the present application is to provide a solar cell appearance defect detection method.

[0006] The present application solves the technical problem by adopting the technical solution as follows:

[0007] A solar cell appearance defect detection method, characterized in that the method comprises the following steps:

[0008] Step 1: Collecting solar cell appearance images, including defect and non-defect images;

[0009] Step 2: Building an appearance defect detection model, the model comprising a multi-scale feature extraction network, a cross-level parallel feature fusion network and a classification regression network;

[0010] The multi-scale feature extraction network is composed of a plurality of multi-scale residual feature extraction modules connected in series, and the feature maps b1, b2 and b3 output by the third-to-last, second and first multi-scale residual feature extraction modules are multi-scale feature maps obtained by the multi-scale feature extraction network; the input of the multi-scale residual feature extraction module is subjected to a segmentation operation by a CBS module to obtain two feature maps X 00 and X 01 ; after the feature map X 00 is subjected to a splitting and merging module, it is spliced with the feature map X 00 to obtain a feature map X1; after the feature map X1 is subjected to a splitting and merging module, it is spliced with the feature map X1 to obtain a feature map X2; the process is repeated k times to obtain a feature map X k ; after the feature map X k is spliced with the feature map X 01 , it is subjected to a splitting and merging module to obtain the output of the multi-scale residual feature extraction module; the cross-level parallel feature fusion network is divided into two stages in the horizontal direction and three parallel branches in the vertical direction, and the information of each branch is fused with each other to make the expression ability of the features stronger;

[0011] Step 3: Training the appearance defect detection model using the images obtained in Step 1, and using the trained appearance defect detection model for appearance defect detection of solar cells.

[0012] Further, the first and second branches of the cross-level parallel feature fusion network each include two HP modules, and the third branch includes one HP module, and the HP modules realize high-precision feature fusion through a combination of convolution and residual connection; the feature map b1 is added to the output of the first HP module of the second branch and the output of the HP module of the third branch after passing through the first HP module of the first branch, and then passes through the second HP module of the first branch to obtain the output of the first branch; the feature map b2 is added to the output of the first HP module of the first branch and the output of the HP module of the third branch after passing through the first HP module of the second branch, and then passes through the second HP module of the second branch to obtain the output of the second branch; the feature map b3 is added to the output of the first HP module of the first branch and the output of the first HP module of the second branch after passing through the HP module of the third branch to obtain the output of the third branch; the outputs of the three branches are the feature maps obtained by the cross-level parallel feature fusion network.

[0013] Further, the HP module includes two serial basic blocks, each basic block includes four serial convolution units, and each convolution unit includes two convolution layers, and the output of the second convolution layer is connected in residual connection with the input of the convolution unit to obtain the output of the convolution unit.

[0014] Further, the input of the splitting and merging module passes through four 1x1 convolution layers to obtain feature maps P1-P4; the feature maps P2, P3 and P4 pass through a convolution layer to generate feature maps Y2, Y3 and Y4 respectively; the feature maps Y1-Y4 pass through splicing and then pass through a convolution operation to obtain the output of the module; the feature map Y1 is directly generated from the feature map P1.

[0015] Further, the CBS module is composed of a convolution layer, a batch normalization layer and an activation layer.

[0016] Compared with the prior art, the present application has the following advantages:

[0017] (1) As the network goes deeper, pooling, convolution and other operations can cause the disappearance of tiny and weak defect features, therefore the application constructs a multi-scale feature extraction network to extract feature maps of different scales, retains defect features of different scales, and avoids the loss of tiny and weak defect information. The multi-scale residual feature extraction module is used to segment the feature map, and the segmented feature map has small size and small receptive field, which can better utilize fine-grained feature information; the output of the SAM module is connected with the input in residual connection, which better combines local information and global context information, improves the global perception performance of the model, and the residual features contain rich shallow spatial information, and the spatial information contains more detailed features, so the detection ability of the model for tiny and weak defects in the appearance of the photovoltaic cell is improved. The SAM module further expands the receptive field, enriches the expression ability of the feature map, and enhances the feature extraction ability of the model; and the SAM module ensures that the output depth does not increase, thereby reducing the calculation amount.

[0018] (2) The cross-level parallel feature fusion network fuses multi-scale features, connects the information flow from high to low in a parallel manner, repeatedly uses the information flow of different branches, and retains as many features in the image as possible, so as to obtain sufficient feature representation. The cross-level parallel feature fusion network combines bidirectional multi-scale features between shallow and deep layers to prevent the loss of tiny and weak defect information in the shallow features in the feature fusion process. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a structural diagram of the multi-scale feature extraction network and the cross-level parallel feature fusion network;

[0020] Figure 2 is a structural diagram of the MRFE module;

[0021] Figure 3 is a structural diagram of the SAM module;

[0022] Figure 4 is a structural diagram of the cross-level parallel feature fusion network;

[0023] Figure 5 is a structural diagram of the HP module;

[0024] Figure 6 is a structural diagram of the basic block. DETAILED DESCRIPTION

[0025] Specific embodiments will be described below with reference to the accompanying drawings, which are only used to specifically illustrate the technical solutions of the application, and do not limit the protection scope of the application.

[0026] The application is a solar cell appearance defect detection method (referred to as method, see Figures 1-6 ), comprising the following steps:

[0027] First step: acquisition and division of dataset

[0028] The appearance images of the solar cell pieces are collected by using an industrial camera, including defect images and non-defect images; the defect images are preprocessed through flipping, rotating, color enhancement and other operations to realize expansion of defect data; the defect images are manually labeled to obtain real labels of defects; in this embodiment, a total of 5763 defect images are obtained, and the defect categories include edge collapse, broken grid, slurry leakage and dirty piece, among which there are 1211 edge collapse images, 1849 broken grid images, 1648 slurry leakage images and 1055 dirty piece images; the training set and the test set are randomly divided according to the ratio of 8:2, and the training set is further divided into the final training set and the validation set according to the ratio of 8:2, that is, the final training set: validation set: test set = 0.64: 0.16: 0.2; the final training set has a total of 3688 images, including 775 edge collapse images, 1183 broken grid images, 1054 slurry leakage images and 675 dirty piece images; the validation set has a total of 922 images, including 194 edge collapse images, 296 broken grid images, 264 slurry leakage images and 169 dirty piece images; the test set has a total of 1653 images, including 242 edge collapse images, 370 broken grid images, 330 slurry leakage images and 211 dirty piece images.

[0029] Second step: construction of appearance defect detection model

[0030] As shown in Figure 1 , the appearance defect detection model includes a multi-scale feature extraction network, a cross-level parallel feature fusion network (Cross-level Parallel Feature Fusion Module, CPF2 for short) and a classification regression network; the multi-scale feature extraction network is used to extract multi-scale features from the input image, the cross-level parallel feature fusion network is used to efficiently fuse the multi-scale features to obtain three feature maps of different scales; and the three feature maps pass through the classification regression network to obtain the detection result.

[0031] The multi-scale feature extraction network is composed of a plurality of (5 in this embodiment) multi-scale residual feature extraction modules (Multi-scale Residual Feature Extraction Module, MRFE for short) connected in series. After passing through each MRFE module, the size of the feature map becomes half of the original size. The structure of the MRFE module is shown in Figure 2 , the input of the MRFE module passes through a CBS module to obtain a feature map X, and the size of the feature map X is C×H×W, C represents the number of channels, H represents the height, and W represents the width; the feature map X is split to obtain two feature maps X 00 and X 01 , and the sizes of the two feature maps are both (C / 2)×H×W; the feature map X 00After a split-and-merge module (SAM), and with feature map X 00 The feature map X1 is obtained by concatenating the features map X1 with the feature map X2 after passing through a SAM module. This process is repeated k times to obtain the feature map X. k Feature map X k With feature map X 01 After concatenation, the image passes through a SAM module to obtain feature map L, which is the output of the MRFE module. The multi-scale feature extraction network outputs three feature maps in total, namely the outputs of the third, second, and third MRFE modules from the end, i.e., feature maps b1, b2, and b3. The resolutions of feature maps b1, b2, and b3 are one-eighth, one-sixteenth, and one-thirty-second of the original input image of the model, respectively. The CBS module consists of a concatenated convolutional layer (Conv), a batch normalization layer (BN), and an activation layer (SILU).

[0032] The formula for calculating the feature map L is as follows:

[0033] L = SAM(X) k (1)

[0034]

[0035] Where SAM(·) represents the operation of the SAM module, concat(·) represents the concatenation operation, and X k-1 X T Both represent feature maps.

[0036] like Figure 3 As shown, the input to the SAM module passes through four 1×1 convolutional layers, dividing the feature map into four identical feature maps P1 to P4 according to channels. The number of channels in feature maps P1 to P4 is one-quarter of the number of channels in the input feature map. Feature maps P2, P3, and P4 are each passed through a convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1, generating feature maps Y2, Y3, and Y4. Feature map Y1 is directly generated from P1. After concatenation, feature maps Y1 to Y4 are passed through a 1×1 convolution to obtain the output of the SAM module. The output feature map of the SAM module has the same size as the input feature map.

[0037] The MRFE module divides the feature map X into two parts. The divided feature map has small size and small receptive field, which can better utilize fine-grained feature information. After convolution, more detailed information is retained. The MRFE module not only has multi-scale feature extraction capability, but also can use the position information of the accurate feature map to construct the relationship between channels and capture long-distance dependency. The splicing operation is used multiple times in the MRFE module to complete the transfer of residual features, better combining local information and global context information, and improving the global perception performance of the model. The residual features contain rich shallow spatial information, and the spatial information contains more detailed features, thus improving the detection capability of the model for small and weak defects in the appearance of the photovoltaic cell. The SAM module further expands the receptive field, enriches the expression ability of the feature map, and enhances the feature extraction capability of the model. Moreover, the SAM module ensures that the output depth does not increase, thereby reducing the amount of calculation.

[0038] As shown in Figure 4 , the cross-stage parallel feature fusion network is composed of multiple branches with different resolutions across multiple stages, and is divided into two stages in the horizontal direction and three parallel branches with different resolutions in the vertical direction. The first and second branches each include two high-precision modules (HP), and the third branch includes one HP module. The feature maps b1, b2 and b3 are the inputs of the three branches. After the feature map b1 passes through the first HP module of the first branch, it is added to the output of the first HP module of the second branch and the output of the HP module of the third branch, and then passes through the second HP module of the first branch to obtain the output of the first branch. After the feature map b2 passes through the first HP module of the second branch, it is added to the output of the first HP module of the first branch and the output of the HP module of the third branch, and then passes through the second HP module of the second branch to obtain the output of the second branch. After the feature map b3 is added to the output of the first HP module of the first branch and the output of the first HP module of the second branch, it passes through the HP module of the third branch to obtain the output of the third branch. The cross-stage parallel feature fusion network obtains three feature maps with different scales, which are the outputs of the three branches. The cross-stage parallel feature fusion network can generate high-precision feature maps through multi-scale fusion, which have higher feature expression capability, better spatial perception and stronger anti-interference capability, thereby helping to detect small and weak defects in the appearance of the photovoltaic cell.

[0039] As shown in Figure 5 , the high-precision module (HP) includes two basic blocks (Basic Block) connected in series. The HP module can fuse information of different branches, strengthen the interaction of information flow, and better detect small and weak defects. Referring to Figure 6The basic block comprises four convolution units connected in series, each convolution unit comprising two convolution layers, and the output of the second convolution layer is added to the input of the convolution unit, that is, a residual connection, to obtain the output of the convolution unit; for each convolution unit, the first convolution layer has a convolution kernel size of 3*3, a step of 1, and a padding of 1; and the second convolution layer has a convolution kernel size of 5*5, a step of 2, and a padding of 1.

[0040] The classification regression part: the three feature maps obtained by the cross-level parallel feature fusion module are respectively subjected to a convolution layer to obtain three feature maps of different scales, that is, the prediction results of the appearance defect detection model, and the prediction results comprise the category, position and confidence of the defect.

[0041] Step 3: training and testing of the appearance defect detection model

[0042] (1) image preprocessing

[0043] The training set obtained in the first step is subjected to data enhancement by Mosic data enhancement, that is, one image is selected as an output image, and then three images are randomly selected, and the four images are cropped and spliced to form a new image;

[0044] (2) setting network parameters and initializing the appearance defect detection model

[0045] The maximum number of iterations (epoch) is set to 200; the learning rate is set to 0.001 for the first 50 epochs, and the learning rate is set to 0.0001 for the last 150 epochs; the learning rate reduction factor is 0.1, and the weight decay of the regular term is 0.0005;

[0046] (3) the preprocessed training set is input into the appearance defect detection model for training, the validation set obtained in the first step is used to verify the trained appearance defect detection model, and the model parameters are optimized through back propagation;

[0047] (4) after the image to be detected is subjected to size normalization, it is input into the trained appearance defect detection model for defect detection to obtain the category, position and confidence of the defect.

[0048] In order to verify the effectiveness of the method, the method is compared with common defect detection models, and the detection results are shown in Table 1.

[0049] Table 1: detection result statistics of different models

[0050]

[0051] From the table, it can be seen that the average detection accuracy of the method for the four defects is higher than that of the remaining common defect detection models, verifying the effectiveness of the method.

[0052] The unmentioned parts of the present application apply to the prior art.

Claims

1. A method for detecting appearance defects in solar cells, characterized in that, The method includes the following steps: Step 1: Acquire images of the solar cell's appearance, including both defective and defect-free images; Step 2: Construct an appearance defect detection model, which includes a multi-scale feature extraction network, a cross-level parallel feature fusion network, and a classification and regression network. The multi-scale feature extraction network consists of multiple multi-scale residual feature extraction modules connected in series. The feature maps b1, b2, and b3 output by the third, second, and third-to-last multi-scale residual feature extraction modules are the multi-scale feature maps obtained by the multi-scale feature extraction network. The input of the multi-scale residual feature extraction module is segmented by a CBS module to obtain two feature maps X. 00 and X 01 Feature map X 00 After a splitting and merging module, and with feature map X 00 The feature map X1 is obtained by concatenation; after passing through a splitting and merging module, feature map X1 is concatenated with feature map X2 to obtain feature map X3; this process is repeated k times to obtain feature map X4. k Feature map X k With feature map X 01 After splicing, the output of the multi-scale residual feature extraction module is obtained through a splitting and merging module. The cross-level parallel feature fusion network is divided into two stages horizontally and three parallel branches vertically. The information of each branch is fused together, making the feature expressive power stronger. The first and second branches of the cross-level parallel feature fusion network each include two HP modules, and the third branch includes one HP module. The HP modules achieve high-precision feature fusion through a combination of convolution and residual connections. Feature map b1, after passing through the first HP module of the first branch, is added to the output of the first HP module of the second branch and the output of the HP module of the third branch, and then passes through the second HP module of the first branch to obtain the output of the first branch. Feature map b2, after passing through the first HP module of the second branch, is added to the output of the first HP module of the first branch and the output of the HP module of the third branch, and then passes through the second HP module of the second branch to obtain the output of the second branch. Feature map b3, after being added to the output of the first HP module of the first branch and the output of the first HP module of the second branch, passes through the HP module of the third branch to obtain the output of the third branch. The outputs of the three branches are the feature maps obtained by the cross-level parallel feature fusion network. Step 3: Use the images obtained in Step 1 to train the appearance defect detection model, and then use the trained appearance defect detection model to detect appearance defects in solar cells.

2. The method for detecting appearance defects in solar cells according to claim 1, characterized in that, The HP module includes two cascaded basic blocks, each basic block includes four cascaded convolutional units, each convolutional unit includes two convolutional layers, and the output of the second convolutional layer is residually concatenated with the input of the convolutional unit to obtain the output of the convolutional unit.

3. The method for detecting appearance defects in solar cells according to any one of claims 1 to 2, characterized in that, The inputs of the splitting and merging modules are passed through four 1×1 convolutional layers to obtain feature maps P1 to P4; feature maps P2, P3, and P4 are each passed through a convolutional layer to generate feature maps Y2, Y3, and Y4; feature maps Y1 to Y4 are concatenated and then subjected to one more convolutional operation to obtain the output of the module; feature map Y1 is directly generated from feature map P1.

4. The method for detecting appearance defects in solar cells according to claim 1, characterized in that, The CBS module consists of convolutional layers, batch normalization layers, and activation layers connected in series.

Images