Research On Defect Detection Method Of Solar Cell Based On Improved Generative Adversarial Network

The solar energy technology has become one of the strategic development goals of various countries,and solar cells are the most important carrier.Currently,monocrystalline solar cells with the highest conversion efficiency are highly susceptible to defects such as breakage,scratches and broken grids during the production process,which seriously affect their service life and photovoltaic conversion efficiency.These defects,which exist inside solar cells,are often coupled in their non-uniformly distributed complex textures,which increases the complexity of detecting them accurately.The collection of solar cell data samples is often accompanied by the problem of light imbalance,which makes the detection accuracy of traditional deep learning methods suffer.Meanwhile,the preparation and labeling process of the preliminary dataset consumes a lot of cost and time,and the inability to collect samples for all types of defects and the difficulty of manual labeling to ensure that each sample is accurate leads to a trained model with a large false detection leakage problem.Therefore,this paper proposes a defect detection method for solar cell images based on generative adversarial network.The specific research contents and research results are as follows:(1)For the solar cell image illumination imbalance problem,a multi-branch network structure image enhancement method with two attention mechanisms is proposed in this paper.To solve the problem of easy blurring or noise amplification in the enhanced image,two attention sub-networks are used to simultaneously guide the model for brightness adjustment and noise removal of the light imbalance images.By designing an end-to-end multi-branch network structure,the feature extraction ability of the model for low-light regions is effectively enhanced,the amount of image information is strengthened,and the focused enhancement of the under-illuminated regions of the image is achieved.In addition,this paper introduces a content loss function to use higher latitude information to enhance the visual effect of the image,draws on the VGG network as a content extractor,and measures the effect of the enhanced image by calculating the difference between the output image of the multibranch network and the output image of the VGG-19 network.The experimental results show that the enhanced images processed by the method in this paper not only can effectively avoid the problem of image distortion,but also the brightness distribution of the image is more balanced and the dynamic range is reduced;compared with other algorithms,the enhancement method proposed in this paper has significant enhancement effects in terms of contrast,sharpness and information of the images.(2)For the insufficient labeling data and complex texture background of solar cells,a multi-scale depth residual-based generative adversarial network MSDR-CycleGAN image reconstruction method based on unsupervised learning is proposed.To solve the problem of complex texture reconstruction of solar cells,the generator is improved based on multi-scale and residual structure,and the improved generator has a better ability to learn the detail information of the input defect image and can distinguish the defects from the complex background in the image more effectively,so as to realize the repair of defect samples of solar cells.The background similarity between the reconstructed image and the original image is constrained by cyclic consistency loss and structural similarity loss to ensure that the background information of the reconstructed image remains unchanged.Finally,three generative adversarial quality evaluation indexes are used to verify the model’s ability to repair defective samples in this paper.The experimental results show that the MSDR-CycleGAN model proposed in this paper has a better reconstruction effect on the defective solar cell images with non-uniform complex texture backgrounds compared with the other image reconstruction methods such as Single-GAN and the original CycleGAN.(3)For the actual defect detection task of solar cell images,the reconstructed repair image is obtained by the generator based on the MSDR-CycleGAN model.By calculating the absolute difference between the reconstructed image and the original image,the location of the defect in the image is determined.The generator network parameters in this model have been pre-optimized so that it can successfully convert the defective image in the source domain to the corresponding defect-free image in the target domain.The information such as texture background is kept highly consistent between the reconstructed image and the original image except for the defective areas.The residual image is computed by finding the absolute difference between the original defect image and the repaired image,and morphological operations such as binarization and threshold segmentation are applied to enhance the defect regions of the image,and the black part of the test result represents the content of the identified defects.The experimental results show that the MSDR-GAN model exhibits high accuracy and reliability with a false detection rate as low as 0.03.Compared with other algorithms such as Single-GAN and the original CycleGAN,it shows high effectiveness and superiority in the solar cell image defect detection task.