A New Feature Pyramid Network With Bidirectional Jump Connection Network for Small Defect Detection on Solar PV Wafer

Photovoltaic (PV) wafers are the core component of solar cells, and detection of their surface defects is critical to improving the reliability of solar cell. The intricacies inherent of wafer production process give rise to numerous small defects on the surface of PV wafers, significantly challenging the task of detecting defects. In this study, feature pyramid network with bidirectional jump (BJFPN)-Yolov5 is proposed as a defect detection network to detect small defects. First, for the issue of limited effective features for small defects, the BJFPN connection is obtained by integrating a bidirectional jump connection into Yolov5's neck, which can alleviate semantic details loss in the upsampling procedure. Second, the channel compression module (CCM) was devised utilizing the channel attention mechanism, and it can enable the elimination of redundant feature information and accentuate crucial details pertaining to small defects. Third, the proposed structures of BJFPN and CCM are integrated into Yolov5 to form BJFPN-Yolov5. The performance of BJFPN-YoLov5 for detecting small and tiny defects has been validated using a real-world PV Wafer dataset, surpassing other state-of-the-art defect detection methods. The BJFPN-Yolov5 is implemented on the Jetson Nano B01 edge device, and the results demonstrate that it can deal with real-time defect detection.