Solar Surface Defect Detection Based on Optimized YOLOv5

Abstract

This study introduces an advanced Solar Cell Surface Defect Detection method utilizing an improved YOLO v5, FaserRCNN and YOLOV6 algorithms. Addressing the challenges posed by complex image backgrounds, variable defect morphology, and large-scale differences, our approach incorporates deformable convolution in the CSP module for adaptive learning scale and perceptual field size. The integration of the ECA-Net attention mechanism enhances feature extraction capabilities, while the addition of a tiny defect prediction head improves detection accuracy across different scales. Optimization techniques, including Mosaic and MixUp data augmentation, K-meansCC clustering anchor box algorithm, and the CIOU loss function, contribute to superior model performance. Experimental results demonstrate an impressive accuracy of 97.14% for YOLOv5, outperforming Faster R-CNN's 90.66%. Further extension studies on YOLOv6, YOLOv7, and YOLOv8 reveal YOLOv6 as the most effective, achieving a remarkable accuracy of 98.28%. This research establishes a robust solution for solar cell defect detection, showcasing the efficacy of our proposed algorithm for industrial applications.