Few-shot PCB defect detection based on meta-feature enhancement
SONG Tao；LI Cheng；XIONG Hailong；YE Dingxing；YUAN Chuan；ZHAO Yuewen；TANG Hongyao；RAN Lu；
A meta-learning approach is introduced for the task of PCB surface defect detection under few-shot conditions，aiming to extract prior knowledge and achieve rapid generalization on novel defects. Simultaneously，a few-shot detection algorithm based on meta-feature enhancement is designed. Firstly， meta-learning is combined with fine-tuning strategy，where only the detector head is fine-tuned during the meta-testing phase，improving the classification ambiguity in the knowledge transfer process. To address the issue of confusion between novel and base class defects on PCB， a global feature fusion module is designed in the support branch to fuse global channel features with original support features，distinguishing different defect categories. Finally，a self-attention module is introduced in the query branch to enhance the network's focus on small targets，helping to solve the problem of defect target omission. The method proposed in this paper demonstrates excellent detection performance in the 10-shot task，achieving a novel class Average Precision（AP）of 62.4% on the PKU-Market-PCB defect dataset.
