PIDDN: Pair-Image-Based Defect Detection Network With Template for PCB Inspection

Abstract

The detection of defects has consistently posed a significant obstacle in the domain of machine vision, particularly in the context of the production of printed circuit board (PCB) components. To advance this technology, we present pair-image-based defect detection network (PIDDN), a novel framework that enables precise detection of object bounding boxes by utilizing a pair of images, with one designated as a template. The PIDDN approach involves the utilization of a Siamese neural network to simultaneously encode the features of the pair image, followed by the implementation of a template feature fusion network (TFFN) for integration. In addition, we introduce a template feature rectification module (TFRM) that aligns the feature maps of the pair-images attentively. We evaluate PIDDN on the DeepPCB dataset, and it achieves an impressive mean average precision (mAP) score of 99.6%. Furthermore, we present PairPCB, a complex and realistic dataset collected from real-world PCB production scenarios, to validate the effectiveness of template images. Extensive experiments demonstrate that PIDDN outperforms mainstream object detection algorithms with a 4.1% improvement in mAP. Code will be available at: https://github.com/QixingJiang/PIDDN.