An Enhancement Generative Adversarial Networks Based on Feature Moving for Solar Panel Defect Identification

Unbalanced dataset is a common problem in the practical application of deep learning algorithms, and it often limits the diagnostic ability of identification models. Therefore, a sample enhancement method combining generative adversarial network (GAN) and feature moving is proposed for solar panel defect detection, thereby improving the defect recognition accuracy. This method is mainly divided into the following three steps: first, the traditional GAN network is used to generate samples to reduce the problem of data imbalance; then, the feature moving method is proposed to be integrated into the YOLOv5S model to reduce the distribution difference between the generated samples and the actual samples, which further improving the sample quality and enhance the accuracy of defect identification; finally, the proposed method is effectively verified by three types of unbalanced solar panel defect data (over-etching, fingerprint, and scratch). When using GAN to generate samples, the results of the ablation and comparison experiment show that the single-class average precision (AP) can be increased by up to 2.8% and the mean AP (mAP) can be increased by 1%. When using styleGAN3 to generate samples, AP and mAP can be improved by 0.2%.