Multi-Granularity Relation Enhancement Network for Tiny Defect Detection on Printed Circuit Board

Tiny defect detection on printed circuit boards (PCBs) is a crucial task to ensure the quality of the downstream electronic products. Despite promising results, prior works often overlook the contextual surroundings of candidate defect proposals and relations among the proposals, hindering their performance. To this end, we propose a Multi-Granularity Relation Enhancement Network (MGRE-Net) to enhance the representation of tiny defects on PCBs by fusing image- and instance-level cross-domain information, which consists of two pivotal mechanisms: the image-to-instance contextual enhancement and the instance-to-instance visual reasoning. Specifically, the former is proposed to reweigh the bias of each instance-level proposal by aggregating image-level contextual features, whereby proposals related to defects are assigned higher biases. Based on reweighed proposals, the latter is further designed to adaptively encode a sparse graphical representation of key high-level semantic relations among proposals for defect reasoning. Extensive experiments are conducted on four industrial datasets, and the results clearly illustrate that MGRE-Net surpasses previous baselines for detecting tiny defects on PCBs.