An Enhanced YOLOv8 Network Combined with Model Pruning for Defect Detection in Solar Cell Images

During the manufacturing and transportation of solar cells, product defects are inevitable. Rapid and accurate defect detection is crucial for improving the quality of photovoltaic panels. This paper proposes a solar cell defect detection model based on YOLOv8, aimed at achieving efficient and highly accurate defect detection in electroluminescence images of solar cells. The High-level Screening Path Aggregation Network enhances the model's capability to filter feature information. The group normalization shared convolution head improves parameter utilization and computational efficiency. Group-level Pruning is employed to remove a significant number of redundant parameters, thereby substantially reducing computational complexity. The proposed model achieves a mAP of 95.9% on the PVEL-AD, with a 62.96% reduction in computational load compared to YOLOv8n. This represents a comprehensive enhancement in detection efficiency and accuracy, demonstrating significant practical value in solar cell defect detection.