Defect Diagnosis of Photovoltaic Module Visible Light Images Under Imbalanced Sample Conditions

To address the issues of high miss and false detection rates in defect detection of PV (photovoltaic) module visible light images under imbalanced data conditions, an improved data augmentation method based on DCGAN (Deep Convolutional Generative Adversarial Networks) is proposed. Firstly, in the DCGAN model, fully connected layers are incorporated into both the generator and the discriminator, and the transpose convolutional layers and convolutional layers are replaced with residual blocks. This enables the generation of $512\times 512$ resolution images using fewer network layers and avoids the problems of gradient vanishing or gradient exploding caused by excessive layers. Secondly, a self-attention mechanism is introduced into both the generator and the discriminator to enhance the ability of DCGAN to extract image features, allowing the model to capture small defects or subtle differences in defects on PV modules more accurately. Finally, to address the issue of DCGAN generating repeated and similar images, an image similarity measurement function is introduced into the loss function. The dataset generated by improved DCGAN achieved mAP scores of 89.3% and 90.9% on YOLOv5 and YOLOv8 models, respectively, showing increases of 0.7% and 1.2% compared to the original models, which validates the effectiveness of the enhancements.