Study on PV Defect Detection Based on CWE YOLOv8

This study proposes an enhanced photovoltaic panel defect detection method based on an optimized YOLOv8 model, addressing common defects like scratches, broken grids, and dirt. To improve feature extraction for small or blurry defects, the C2f-DySnakeConv module is added to the backbone network. The traditional upsampling process is replaced by the WaveletPool module, which uses wavelet transforms to enhance multi-scale features and improve feature resolution. A lightweight EfficientHead detection head is also introduced to boost the model’s accuracy and robustness in complex detection scenarios. Additionally, to combat sample imbalance and improve the classification of challenging defects, a dynamic weight-adjusted Focal Loss function is employed. Experimental results show that the proposed CWE-YOLOv8 model achieves an mAP of 87.9%, a 5.1% improvement over the YOLOv8n baseline. Despite these advancements, the model retains a compact size of 7.26MB, only slightly larger than the original model’s 5.94MB. This approach offers a high-precision, real-time, and lightweight solution for PV defect detection, with notable potential for engineering applications.