A Comparative Study of Deep Learning Architectures for Photovoltaic Panel Defect Detection

This study assesses the performance of four deep learning models—AlexNet, ResNet50, VGG16, and Vision Transformer (ViT)—for classifying defects in solar panels, including bird droppings, dust, snow cover, and physical or electrical damage. Trained and tested on a dataset of 891 images, the models were evaluated for accuracy, computational efficiency, and interpretability. The findings reveal the strengths and limitations of each model, with ViT demonstrating superior ability to capture complex image patterns through its self-attention mechanism. This suggests that transformer-based models, particularly ViT, offer significant advantages for detecting defects in solar panels.