A YOLO-based defect detection system for printed circuit boards

An improved YOLOv5 for PCB defect detection named as YOLO-PDD is proposed in this paper. Based on this, a circuit board defect detection system which can achieve high accuracy and fast detection of small target defects under complex background noise interference is designed. The industrial cameras and LED light sources are used in the system to capture images of circuit board defects, forming a self-made original dataset. At the same time, stitching technology is adopted to enhance the original dataset. YOLO-PDD has made the following improvements to the original YOLOv5 algorithm: Firstly, a double-layer routing attention module BRA is introduced in the feature extraction network. Next, the backbone of YOLOv5 algorithm is replaced by DenseNet. Finally, a feature pyramid network which is bidirectional and weighted in nature called Bifpn is added to the YOLOv5 feature fusion network. The circuit board defect detection system software is built based on QT and can obtain real-time detection results and data statistics online. Experimental results show that the system can accurately detect defect targets in circuit board images at a speed of 47ms per image, achieving 92.28% mAP on our self-made dataset, which is superior to the average detection accuracy of current state-of-art original detection algorithms and has good practicality and effectiveness.