This study shows a new way to find flaws on the surface of a solar cell using better versions of the YOLO v5, FaserRCNN, and YOLOV6 algorithms. Our method includes flexible convolution in the CSP module for adjustable learning scale and visual field size. This helps with problems like complex picture backgrounds, flaw shape that changes, and big differences. The addition of the ECA-Net focus method improves the ability to extract features, and the addition of a small flaw prediction head improves the accuracy of spotting at different scales. A better model `works with optimization methods like the CIOU loss function, K-meansCC clustering anchor box algorithm, Mosaic and MixUp data enrichment, and more. The experimental results show that YOLOv5 is much more accurate than Faster R-CNN, RPN which only achieved 90.66% accuracy. Additional extension studies on YOLOv6, YOLOv7, and YOLOv8 show that YOLOv6 is the most accurate, with a score of 98.28%. This study creates a strong way to find problems in solar cells and shows that our suggested method works well for business uses.