Detection of Surface Defects in Solar Cells by Bidirectional-Path Feature Pyramid Group-Wise Attention Detector

Due to the multiscale characteristics of defects and strong background interference, the automation of solar cell surface defect detection is still a challenge. To address this problem, this article proposes a novel defect object detector called bidirectional-path group-wise attention detector (BPGA-Detector) which consists of two parts: the bidirectional-path feature pyramid network (BPFPN) and the group-wise attention module (GAM). The BPFPN combines multiscale features using a bidirectional-path feature fusion method that structured by connecting the bottom-up path feature pyramid network (FPN) to the original FPN, preserving the characteristics of minor and weak flaws in the shallow layer. Furthermore, the GAM is elaborately designed to suppress the background disturbance and highlight the defect locations by connecting multilayer contextual features, which significantly improves the discriminant ability of small defects. Finally, the experimental results on a largescale solar cell dataset including 6263 images, 5763 of which are defective, demonstrate that the proposed method achieve superior detection performance (mAP50 up to 88.8%).