Crystalline silicon photovoltaic cell defect real-time detection method based on multi-scale feature fusion

Abstract

The invention relates to a crystalline silicon photovoltaic cell defect real-time detection method based on multi-scale feature fusion. The method comprises the following steps: establishing a crystalline silicon photovoltaic cell defect electroluminescent image detection data set; constructing a photovoltaic cell defect real-time target detection model based on RT-DETR improvement; training the photovoltaic cell defect target detection model by using the training set of the detection data set; and detecting a defect image to be detected by using the trained photovoltaic cell defect target detection model. According to the method, the model for detecting the defects of the crystalline silicon photovoltaic cell in real time is provided based on the RT-DETR model, the method does not depend on anchor frame setting and non-maximum suppression processing, and the detection speed is increased; an inverted residual error expansion network IRD-Net serves as a backbone network for extracting image features, the static and dynamic feature modeling capability is enhanced, features based on receptive fields of different scales are extracted, and the detection performance is improved; an extended wavelet cross-scale feature fusion module DWCSFM is introduced to optimize a neck encoder, high-level and low-level features are alternately updated, the number of parameters and the calculation load are reduced, and the detection capacity of the model for defect targets of different scales is further enhanced.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning, computer vision and target detection technology, and in particular to a real-time detection method for crystalline silicon photovoltaic cell defects based on multi-scale feature fusion. Background Art

[0002] In the past decade, solar photovoltaics have received widespread attention as a green energy source worldwide and is currently considered one of the most viable renewable energy sources in the world. The integration of photovoltaic systems with buildings can utilize the advantages of photovoltaic modules to replace some building materials, achieve power generation and reduce building energy consumption. Photovoltaic systems deployed in outdoor environments are subject to a variety of climate stress factors, including solar radiation, humidity, precipitation, wind, snow and dust accumulation. These factors gradually degrade the performance of photovoltaic modules throughout their service life. At the same time, manufacturing, transportation, installation and operation processes may introduce defects such as black cores and microcracks. In addition, the compact distribution and configuration of photovoltaic modules in building-integrated photovoltaic systems place higher requirements on power generation quality, efficiency and fire hazard prediction. Therefore, accurate detection of photovoltaic cell defects is essential to ensure the efficient and safe operation of photovoltaic power generation systems.

[0003] Defect detection methods for crystalline silicon photovoltaic cells can be divided into three categories: photovoltaic cell defect detection methods based on traditional methods, photovoltaic cell defect detection methods based on feature extraction and machine learning, and photovoltaic cell defect detection methods based on deep learning.

[0004] Traditional photovoltaic cell defect detection methods can be further divided into: manual detection method, physical detection method, visual imaging and image processing method. Manual detection is that the inspectors check the defects on the surface of photovoltaic panels one by one, but it will consume more resources, and defects may be missed due to human eye fatigue, resulting in low detection efficiency. Physical detection methods are mostly contact detection algorithms. During the detection process, the cells may be damaged again due to the operator's operating errors; and the physical detection method is only suitable for online monitoring in the manufacturing stage, which has limitations in actual application scenarios. Visual imaging and image processing methods use infrared thermal imaging, photoluminescence imaging, and electroluminescence imaging technologies to obtain photovoltaic cell images, and then use traditional image processing methods to calculate gradient features to highlight image edges and texture features. Although visual imaging and image processing methods avoid direct contact with photovoltaic cells, they have the problems of low detection efficiency and too high requirements for the image acquisition environment; and this method can only solve the binary detection problem of whether there are defects.

[0005] The application of methods based on feature extraction and machine learning in photovoltaic cell defect analysis is becoming increasingly prominent. Many studies have adopted methods that combine electroluminescence imaging with machine learning, such as support vector machines. Although the use of feature extraction and machine learning methods can overcome the problems of high labor costs and low detection efficiency compared to traditional photovoltaic cell defect detection methods, it is difficult to handle various defects that may occur in actual production, and it depends on the quality of manual feature extraction, making it difficult to extract high-level semantic features.

[0006] Photovoltaic cell electroluminescence image defect detection methods based on deep learning have become an effective solution to improve the quality and efficiency of photovoltaic cell defect detection. Convolutional neural network (CNN) is a typical representative of deep learning, which can automatically extract important image features and perform pattern recognition without the need for specific domain knowledge. Current object detection algorithms include two-stage networks represented by Faster R-CNN and single-stage networks represented by SDD, RetinaNet and YOLO. However, the limited receptive field of CNNs weakens their ability to capture global features and long-range dependencies. In traditional object detection frameworks using CNNs, models usually rely on predefined anchor boxes or region proposals. These manually designed components are initial guesses of the location and size of the object, which help the network perform accurate localization and classification tasks. Non-maximum suppression (NMS) and other post-processing techniques introduce additional complexity and have a significant impact on model performance. The emergence of end-to-end detectors based on Transformer, such as DEtection TRansformer (DETR), simplifies the object detection process and achieves remarkable performance. However, in the current research on photovoltaic cell defect target detection, the following problems still exist:

[0007] Problem 1: Computational cost and efficiency. Although the detection accuracy is improved, its high computational cost and long detection time affect the scalability of practical applications. More efficient algorithms are needed to achieve real-time detection with limited computing resources.

[0008] Problem 2: Low accuracy in detecting small target defects. Existing methods perform poorly in detecting small target defects against the complex texture background of electroluminescent images. The reason is that the characteristics of such defects are easily obscured by complex textures.

[0009] Question 3: Balance between accuracy and efficiency: There is still room for improvement in the balance between accuracy and inference speed. Summary of the invention

[0010] The purpose of the present invention is to solve the above-mentioned technical problems and provide a real-time detection method for crystalline silicon photovoltaic cell defects based on multi-scale feature fusion.

[0011] The present invention is achieved by at least one of the following technical solutions.

[0012] A real-time defect detection method for crystalline silicon photovoltaic cells based on multi-scale feature fusion specifically comprises the following steps:

[0013] S1: Establish a dataset for defect electroluminescence image detection of crystalline silicon photovoltaic cells;

[0014] S2: Construct a real-time target detection model for photovoltaic cell defects based on the improved RT-DETR;

[0015] S3: Use the training set of the detection data set to train the photovoltaic cell defect target detection model;

[0016] S4: Use the trained photovoltaic cell defect target detection model to detect the defect image to be detected.

[0017] Furthermore, by acquiring a publicly available crystalline silicon photovoltaic cell defect electroluminescent image dataset, the target detection and annotation tool LabelImg is used to annotate the defect targets of the image, and the collected image data is formatted to form a crystalline silicon photovoltaic cell defect electroluminescent image detection dataset.

[0018] Furthermore, the improved photovoltaic cell defect real-time target detection model based on RT-DETR mainly consists of three parts: an inverted residual dilation network IRD-Net for feature extraction, an enhanced encoder, and a decoder and Head for result prediction.

[0019] Furthermore, the inverted residual dilation network IRD-Net for feature extraction specifically includes a convolution block, a pooling layer, an inverted residual moving block and an dilated reparameterized residual block; convolution block 1 uses a 3×3 convolution with a step size of 2 to reduce the feature map size while retaining spatial information, convolution block 2 uses a 3×3 convolution with a step size of 1 to extract detail features, and uses a pooling layer to aggregate contextual information; the inverted residual moving block uses a 1×1 convolution, an extended window multi-head self-attention mechanism and a depth-separable convolution; the dilated reparameterized residual block automatically extracts multi-scale receptive field features, thereby enabling accurate identification and positioning of photovoltaic cell defects.

[0020] Furthermore, the dilated reparameterized residual block adopts a residual design method; first, the input is processed by two 1×1 convolutions respectively, and their outputs are then processed through different processing paths: path 1 and path 2; in path 1, the output is directly retained; in path 2, the regional residual is processed by a 3×3 convolution; then, multi-rate dilated convolution is used through a branch architecture to perform morphological filtering on features of different receptive field sizes; in the dilated reparameterized blocks of branches 2 and 3 involved in path 2, multiple convolution layers and batch normalization layers are fused into a single convolution layer to increase the inference speed; the non-dilated large convolution kernel is decomposed into small convolution kernels and multiple dilated small convolution kernel layers, which enhances the ability to capture spatial sparse features and improves the perception of complex features; finally, these features are fused through point-by-point convolution to form the final residual;

[0021] Adding the fused features obtained from path 2 to the output features of path 1 can better preserve the detailed information of the original features while adding broader semantic context, thereby improving the overall performance of the network.

[0022] Further, the enhanced encoder includes an intra-scale feature interaction module AIFI and a dilated wavelet cross-scale feature fusion module DWCSFM;

[0023] The dilated wavelet cross-scale feature fusion module DWCSFM consists of a convolution block, a dilated fusion module and wavelet pooling; the dilated wavelet cross-scale feature fusion module performs progressive feature fusion on S3 and S4 obtained from the feature extraction network and the fifth-level feature F5 processed by the AIFI module, which helps to reduce the semantic gap between features at different levels and enhance the feature fusion effect by using the correlation between feature maps at different pyramid levels.

[0024] Furthermore, the dilated fusion module uses multiple convolutional layers and a dilated reparameterization block to enhance the diversity of feature extraction and expand the receptive field; the wavelet pooling is used as a downsampling and upsampling operation of the feature map in the dilated wavelet cross-scale feature fusion module.

[0025] Furthermore, the photovoltaic cell defect target detection model training process is: setting the hyperparameters of the model training process, inputting the training set into the photovoltaic cell defect target detection model, performing online image data enhancement, feature extraction and fusion, outputting predicted values ​​and calculating losses, optimizing the model, and saving the model weights after the model is iteratively trained to convergence.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention is based on the RT-DETR target detection model. The backbone network uses an inverted residual dilation network, introduces a new type of dilated re-parameterized residual block, and adopts a multi-branch architecture and multi-rate dilated convolution. It can extract features of different receptive field sizes without the need for complex predefined anchor frame settings, thereby enhancing the network's detection capability for defect targets of different sizes and improving detection accuracy. Combined with the inverted residual moving block, static and dynamic feature modeling is enhanced, and the inference speed is improved. The encoder part uses a dilated wavelet cross-scale feature fusion module, which suppresses aliasing by alternately updating high-level and low-level features, reduces the number of parameters and computational load, and better integrates local and global semantic information, significantly improving the overall detection performance, and achieving better detection effects while meeting real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A schematic flow chart of a method for real-time defect detection of crystalline silicon photovoltaic cells based on multi-scale feature fusion according to an embodiment of the present invention;

[0029] Figure 2 A structural diagram of the photovoltaic cell defect real-time target detection model based on RT-DETR improvement provided by the present invention;

[0030] Figure 3 This is a structural diagram of the inverted residual dilation network IRD-Net described in the present invention;

[0031] Figure 4 This is a structural diagram of the expanded re-parameterized residual block DRRB according to the present invention;

[0032] Figure 5 This is the structural diagram of the dilated wavelet cross-scale feature fusion module DWCSFM described in the present invention. DETAILED DESCRIPTION

[0033] The following is a detailed description of a real-time defect detection method for crystalline silicon photovoltaic cells based on multi-scale feature fusion according to the present invention in conjunction with the accompanying drawings and embodiments. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies implemented based on the content of the present invention belong to the scope of the present invention.

[0034] like Figure 1 As shown, the present invention provides a real-time detection method for crystalline silicon photovoltaic cell defects based on multi-scale feature fusion, which specifically includes the following steps:

[0035] S1: Obtain a dataset of defective electroluminescent image detection of crystalline silicon photovoltaic cells, annotate the defective areas of the dataset images, and organize the data: modify the acquired images to a unified 640×640 resolution, convert the corresponding bounding box coordinate annotations of the images after the modified resolution, and uniformly modify the annotation format to the YOLO data format to complete the construction of the dataset of defective electroluminescent image detection of crystalline silicon photovoltaic cells.

[0036] As an embodiment, a crystalline silicon photovoltaic cell defect electroluminescence image detection dataset includes open source datasets ELPV and PVEL-AD.

[0037] S2: Design a real-time target detection model for photovoltaic cell defects based on the improved RT-DETR. Its structure is shown in the attached figure. Figure 2 As shown in the figure, the improvement process includes: in the RT-DETR framework, the feature extraction backbone network is replaced with the inverted residual dilation network IRD-Net, and the CCFM module in the encoder is replaced with the dilated wavelet cross-scale feature fusion module DWCSFM; the last feature map S5 of IRD-Net is input into the enhanced encoder. In the enhanced encoder, S5 passes through the intra-scale feature interaction AIFI module, which enhances the intra-scale and inter-scale feature interactions by performing a self-attention mechanism while reducing the computational cost. The resulting feature F5 is then fed into the dilated wavelet cross-scale fusion module DWCSFM to fuse features of adjacent scales. The intersection-over-union-aware query selection module selects high-quality initial target queries from the encoded features to provide a better starting point for the decoder. The encoded features are then passed to the decoder component; a feedforward network generates predicted bounding boxes and category probabilities as outputs.

[0038] The feature extraction backbone network is an inverted residual dilation network, which contains three key components: convolutional blocks, inverted residual shift blocks, and dilated re-parameterized residual blocks DRRB. Figure 3 As shown in the figure. The convolution block uses a 3×3 convolution with a stride of 2 to reduce the feature map size while retaining spatial information, followed by a 3×3 convolution with a stride of 1 to extract detailed features, and a pooling layer to aggregate contextual information. The inverted residual mobile block iRMB combines the static local modeling efficiency of CNN and the dynamic global modeling capability of Transformer, using 1×1 convolution, extended window multi-head self-attention mechanism (EW-MHSA) and depthwise separable convolution (DW-Conv). This expands the receptive field of the model and enhances its performance in downstream detection tasks. The inverted residual dilation network uses a dilated reparameterized residual block to automatically extract multi-scale receptive field features, enabling accurate identification and location of photovoltaic cell defects.

[0039] The dilated re-parameterized residual block DRRB structure is shown in the attached Figure 4 As shown in Figure 1. First, the input is processed by two 1×1 convolutions, and their outputs are then processed through different processing paths: path 1 and path 2. In path 1, the output is directly retained. In path 2, the regional residual is processed by a 3×3 convolution. Subsequently, morphological filtering is performed on features of different receptive field sizes using multi-fold dilated convolutions through the branch architecture to obtain features under different receptive fields, so that the feature map has diversity at different scales, thereby improving the network's perception of objects of different sizes. In the dilated reparameterization blocks of branches 2 and 3, multiple convolutional layers and batch normalization layers are fused into a single convolutional layer to increase the inference speed. The non-dilated large convolution kernel is decomposed into small convolution kernels and multiple dilated small convolution kernel layers, which enhances the ability to capture spatial sparse features and improves the perception of complex patterns, optimizing the model's ability to efficiently process complex spatial relationships. For branch 2, K=5 is set, where the convolution kernel size of the parallel convolution layer is k=(3,3) and the dilation rate is r=(1,2). When performing inference in branch 2, the three batch normalization-convolutional layers are reparameterized into a single 5×5 convolution; for branch 3, K=7 is set, where the convolution kernel size of the parallel convolutional layer is k=(5,3,3) and the dilation rate is r=(1,2,3). When performing inference in branch 3, the four batch normalization-convolutional layers are reparameterized into a single 7×7 convolution. Finally, these features are fused through point-by-point convolution to form the final residual. Adding the fused features obtained from path 2 to the output features of path 1 can better preserve the detailed information of the original features while adding a wider semantic context, thereby improving the overall performance of the network.

[0040] The enhanced encoder includes the intra-scale feature interaction module AIFI and the dilated wavelet cross-scale feature fusion module DWCSFM. The purpose of introducing DWCSFM is to strengthen the feature fusion process, which is crucial to the RT-DETR encoder architecture, by effectively utilizing the correlation between feature maps at different pyramid levels. The dilated wavelet cross-scale feature fusion module DWCSFM consists of convolution blocks, dilated fusion modules, and wavelet pooling. Its structure is shown in the attached figure. Figure 5 The dilated wavelet cross-scale feature fusion module performs progressive feature fusion on S3 and S4 obtained from the feature extraction network and the fifth-level feature F5 processed by the AIFI module, which helps to reduce the semantic gap between features at different levels and enhance the feature fusion effect by using the correlation between feature maps at different pyramid levels.

[0041] Specifically, for the convolution block in the dilated wavelet cross-scale feature fusion module, F5 first passes through the convolution block and then is upsampled. The convolution block is calculated as shown in formula (1).

[0042] ConvN = SiLU(BatchNorm(Conv 1×1 (x i )))(1)

[0043] In the formula, SiLU is the Sigmoid gated linear unit activation function, BatchNorm represents batch normalization, and Conv 1×1 represents a convolutional layer with a convolution kernel of 1×1, x i Represents the input features.

[0044] For the dilated fusion module in the dilated wavelet cross-scale feature fusion module, this module aims to enhance the diversity of feature extraction and expand the receptive field by using multiple convolutional layers and a dilated reparameterization block. By adopting convolution kernels with different dilation rates, this module is able to capture features of different scales, thereby covering a larger receptive field without incurring additional computational costs. Smaller dilation rates tend to focus on local detailed information, while larger dilation rates help capture more global semantic information. By fusing these contextual features of different scales, the model is able to better handle defects of different sizes and shapes, thereby improving the accuracy of photovoltaic cell defect detection. In addition, implementing optimized dilated convolution reparameterization has advantages in parameter and computational efficiency, thereby reducing resource consumption while maintaining performance. Conv 1×1 The output of and the upsampling result of F5 are combined with S4 and input into the expansion fusion module to obtain the final feature representation, as shown in formula (2).

[0045] Dilated Fusion=Flatten(Conv 1×1 (cat(x i1 ,x i2 ))+DRB(Conv 1×1 (cat(x i1 ,x i2 )))) (2)

[0046] In the formula, DRB represents the dilated reparameterized block, x i1 and x i2 Represents different input features.

[0047] For the wavelet pooling in the dilated wavelet cross-scale feature fusion module, in order to better capture the semantic relationship in the image, it is used as a downsampling and upsampling operation of the feature map. This method involves decomposing the features into two-level components and discarding the first-level subband, thereby reducing the feature dimension.

[0048] Wavelet downsampling can retain more low-frequency components, retain key information in the feature fusion process and improve the representation ability of the model. By decomposing and selectively retaining the low-frequency part of the signal, wavelet downsampling can effectively reduce the impact of high-frequency noise. Reducing noise can improve feature quality and thus improve the performance of the entire model. In addition, since wavelet downsampling retains more global and semantic information, it effectively improves the quality of fused high-level features. The specific wavelet downsampling operation is shown in formula (3).

[0049]

[0050] In the formula, let X be the input feature map (N, C, H, W), N is the batch size, C is the number of channels, H is the height, and W is the width. Assume that the input feature map X contains four sub-bands: (X LL ,X LH ,X HL ,X HH ). We define four sub-band filters: low-frequency-low-frequency (LL), low-frequency-high-frequency (LH), high-frequency-low-frequency (HL), and high-frequency-high-frequency (HH). We use the four filters to convolve each channel separately, and then concatenate the results by channel to get the output Y.

[0051] Wavelet upsampling enhances feature fusion by using inverse wavelet transform. Through inverse wavelet transform, the wavelet upsampling process can reconstruct the lost high-frequency information. Wavelet upsampling can retain more original spatial structure information, so that the feature map will not lose too many useful spatial details during the upsampling process. Reconstructing images by multi-resolution wavelet coefficients can better retain and restore high-level semantic information. This helps the feature fusion module better combine low-level detail information and high-level semantic information, improving the performance of the overall model. The wavelet upsampling operation is specifically shown in formula (4).

[0052]

[0053] In the formula, the feature map size of the output result Y is (N, C, H, W), and the values ​​of low frequency-low frequency (LL), low frequency-high frequency (LH), high frequency-low frequency (HL), and high frequency-high frequency (HH) are consistent with formula (3).

[0054] S3: Use the training set of the detection data set to train the photovoltaic cell defect target detection model. First, configure the model training hyperparameters, including the size of the input image, the number of defect categories to be detected, the initial learning rate and learning rate optimization strategy, and the number of training rounds. Then input the training data into the model for online image enhancement, feature extraction, feature fusion, output prediction value, calculation loss, and back propagation optimization model. After the network model has been fully trained and converged, save the model's weight file to obtain the photovoltaic cell defect target detection model that has completed training.

[0055] S4: Use the trained photovoltaic cell defect target detection model to detect the defect image to be detected.

Claims

1. A real-time defect detection method for crystalline silicon photovoltaic cells based on multi-scale feature fusion, characterized in that: The specific steps include: S1: Establish a dataset for defect electroluminescence image detection of crystalline silicon photovoltaic cells; S2: Construct a real-time target detection model for photovoltaic cell defects based on the improved RT-DETR; S3: Use the training set of the detection data set to train the photovoltaic cell defect target detection model; S4: Use the trained photovoltaic cell defect target detection model to detect the defect image to be detected.

2. According to claim 1, a real-time detection method for crystalline silicon photovoltaic cell defects based on multi-scale feature fusion is characterized in that: By acquiring the publicly available crystalline silicon photovoltaic cell defect electroluminescence image dataset, the target detection and annotation tool LabelImg is used to annotate the defect targets in the image, and the collected image data is formatted to form a crystalline silicon photovoltaic cell defect electroluminescence image detection dataset.

3. The method for real-time defect detection of crystalline silicon photovoltaic cells based on multi-scale feature fusion according to claim 1, characterized in that: The photovoltaic cell defect real-time target detection model improved based on RT-DETR mainly consists of three parts: an inverted residual dilation network IRD-Net for feature extraction, an enhanced encoder, and a decoder and Head for result prediction.

4. The method for real-time defect detection of crystalline silicon photovoltaic cells based on multi-scale feature fusion according to claim 3, characterized in that: The inverted residual dilation network IRD-Net for feature extraction specifically includes a convolution block, a pooling layer, an inverted residual moving block and an dilated reparameterized residual block; convolution block 1 uses a 3×3 convolution with a step size of 2 to reduce the feature map size while retaining spatial information, convolution block 2 uses a 3×3 convolution with a step size of 1 to extract detail features, and uses a pooling layer to aggregate context information; the inverted residual moving block uses a 1×1 convolution, an extended window multi-head self-attention mechanism and a depth-separable convolution; The dilated reparameterized residual block automatically extracts multi-scale receptive field features, which enables accurate identification and location of photovoltaic cell defects.

5. The method for real-time defect detection of crystalline silicon photovoltaic cells based on multi-scale feature fusion according to claim 4, characterized in that: The dilated reparameterized residual block adopts a residual design method; first, the input is processed by two 1×1 convolutions respectively, and their outputs are then processed through different processing paths: path 1 and path 2; in path 1, the output is directly retained; in path 2, the regional residual is processed by a 3×3 convolution; then, multi-rate dilated convolution is used through a branch architecture to perform morphological filtering on features of different receptive field sizes; in the dilated reparameterized blocks of branches 2 and 3 involved in path 2, multiple convolution layers and batch normalization layers are fused into a single convolution layer to increase the inference speed; the non-dilated large convolution kernel is decomposed into small convolution kernels and multiple dilated small convolution kernel layers, which enhances the ability to capture spatial sparse features and improves the perception of complex features; finally, these features are fused through point-by-point convolution to form the final residual; Adding the fused features obtained from path 2 to the output features of path 1 can better preserve the detailed information of the original features while adding broader semantic context, thereby improving the overall performance of the network.

6. The method for real-time defect detection of crystalline silicon photovoltaic cells based on multi-scale feature fusion according to claim 3, characterized in that: The enhanced encoder includes an intra-scale feature interaction module AIFI and a dilated wavelet cross-scale feature fusion module DWCSFM; The dilated wavelet cross-scale feature fusion module DWCSFM consists of a convolution block, a dilated fusion module and wavelet pooling; the dilated wavelet cross-scale feature fusion module performs progressive feature fusion on S3 and S4 obtained from the feature extraction network and the fifth-level feature F5 processed by the AIFI module, which helps to reduce the semantic gap between features at different levels and enhance the feature fusion effect by using the correlation between feature maps at different pyramid levels.

7. A real-time detection method for crystalline silicon photovoltaic cell defects based on multi-scale feature fusion according to claim 6, characterized in that: The dilated fusion module uses multiple convolutional layers and a dilated reparameterization block to enhance the diversity of feature extraction and expand the receptive field; the wavelet pooling is used as a downsampling and upsampling operation of the feature map in the dilated wavelet cross-scale feature fusion module.

8. The method for real-time defect detection of crystalline silicon photovoltaic cells based on multi-scale feature fusion according to claim 3, characterized in that: The photovoltaic cell defect target detection model training process is as follows: setting the hyperparameters of the model training process, inputting the training set into the photovoltaic cell defect target detection model, performing online image data enhancement, feature extraction and fusion, outputting predicted values ​​and calculating losses, optimizing the model, and saving the model weights after the model is iteratively trained to convergence.

Images