Data Augmentation And Defect Detection Of Solar Cells Based On Generative Adversarial Network

The solar cell is the manufacturer of new clean energy,the defects of which greatly affect the service life and conversion efficiency.The solar cell defect detection technology based on machine vision provides an effective guarantee for the quality of solar cell products.However,due to insufficient solar cell defect sample data and non-uniform random texture distribution in the background,there are many problems when deep learning technologies relied on a large number of training samples and background reconstruction-based methods are applied into solar cell defect detection.Therefore,the solar cell defect data enhancement method based on a generative adversarial network and the non-uniform random texture background reconstruction and defect detection method are proposed.The specific research contents and achievements are as follows:(1)For the insufficient sample of solar cell EL defects and unbalance sample types affecting the performance of defect detection,the normal sample-guided generation adversarial network(NSGGAN)model is proposed to data enhancement of solar cell defect samples.A large amount of normal samples are introduced in the GAN and the loss items are added to promote the expression of defect sample characteristics and improve the diversity of generated samples.In addition,an adaptive weight constraint method is designed to balance the learning ability between generator and discriminator as well as improve the quality of generated samples.In order to verify the effectiveness of the proposed method,this paper verifiers on the solar cells EL dataset,steel strip surface defects dataset and DAGM2007 public dataset.The experimental results show that the MMD,1-NN and other indexes of the images generated by the proposed method are better and higher-quality than those produced through DCGAN and WGAN-GP.Moreover,when the samples enhanced by the proposed data enhancement method are used for classification,obtaining F1 detection indicators are superior to the DCGAN and WGAN-GP models for data enhancement classification results.(2)In order to solve the problems of poor background reconstruction effect and low accuracy of defect detection for EL images of solar cell with non-uniform random texture distribution,the background reconstruction and defect detection method based on counter example generative adversarial network(CEGAN)is proposed.The counter-example samples of the background samples are added to GAN so as to solve the background reconstruction of non-uniform random texture,and the mapping loss term is added to measure the background loss in defect samples.Meanwhile the reconstruction loss term is designed to improve the reconstruction effect of the model on background detail.Finally,by measuring the structural similarity between the input sample and the reconstructed sample,and setting a threshold to distinguish between the defective sample and the normal sample.The experimental results show that compared with the experimental results of background reconstruction using AAE,Cycle GAN and f-Ano GAN,the method proposed in the paper has less background reconstruction error and a higher defect detection accuracy of the solar cell EL images with non-uniform texture background.