An Improved PCB Defect Detector Based on Feature Pyramid Networks

Abstract

Aiming at the problem of false alarm in PCB defect detection in the automatic optical inspection process, many researchers have proposed their methods, but most of them only classify the single defect in single image, and there are multiple defects and multiple categories in single image. In this paper, a real PCB data set consisting of 1540 images generated by AOI is introduced for the detection and classification task. In addition, we propose an improved PCB defect detector based on feature pyramid networks. The detector combines Faster R-CNN and FPN as the infrastructure, and has been adjusted and improved, mainly including the following three innovations: 1) SE module is inserted into the feature extraction backbone network resnet-101, which improves the expression ability of network. 2) An enhanced bottom-up structure is introduced to enhance the whole feature level by using accurate low-level positioning signals. 3) ROI Align is used instead of RoI Pooling to reduce the impact of dislocation on small object defect detection. The experimental results show that, compared with the mainstream object detection network, the proposed method achieves better accuracy, reaching 96.3% mAP, and has better performance for defect detection and classification.