Laser-based automated optical inspection for edge small defect detection in photovoltaic silicon wafers with complex backgrounds

Abstract

Defect detection plays a critical role in ensuring the efficiency of photovoltaic (PV) production lines. Although existing lightweight methods perform well on obvious defects, they struggle with detecting blurred contours and small edge defects in complex backgrounds. This study enhances the YOLOv8 framework by introducing the C2f-WTConv module, which replaces the original C2f block and improves the ability to capture blurred features while reducing the number of parameters by 15.3 %. Additionally, an Efficient Multi-scale Attention (EMA) mechanism is embedded in the neck network to reduce missed detections of small edge defects. The InnerMPDIoU loss function is employed to balance recognition deviations of features and enhance generalization. On the custom SPV-2338 dataset, the proposed YOLOv8-WEIM achieves a mean Average Precision (mAP50) of 83.8 %, representing a 3.3 % increase over the baseline model. Accuracy and recall are improved by 3.1 % and 1.2 %, respectively, while maintaining a frame rate of 118 FPS. Tests on the NEU-DET public dataset further verify the model's generalization capability. The optimized model meets industrial requirements for both speed and accuracy.