Generation of Solar Cell Defect Images Based on Multi-Perceptual Fields and Attention Mechanism

Aiming at the problem of the small sample size of some defect images in solar cell images, a defect images generation model combining multi-perceptual fields and attention mechanism is proposed. Firstly, a generative adversarial network model with dual discriminators is constructed to improve the quality of the generated images; secondly, for focusing accurately on and extracting defect information of different scales with the interference of complex backgrounds, an improved spatial attention module is integrated into the proposed multi-perceptual fields feature extraction and is used in the generator and discriminators; finally, to enhance the texture detail and clarity of the final generated images, the structural similarity (SSIM) loss and the peak signal-to-noise ratio (PSNR) loss are added to the loss function to train the generator, and the generated defect images are mean filtered. The results of the generation experiments on the solar electroluminescence (EL) dataset for 3 different scales of defect images show that the SSIM and PSNR values are higher compared to the existing optimal generation algorithms.