Enhanced YOLO with coordinate attention and BiFPN for printed circuit board defect detection

Abstract

Printed circuit boards (PCBs) are essential components in electronic systems, providing mechanical and electrical support for various integrated elements. Due to their intricate designs, PCBs are prone to manufacturing defects such as short circuits, spurs, and unintended copper formations, which can negatively impact device functionality and reliability. This paper proposes an enhanced You Only Look Once version 8 (YOLOv8) model for detecting and localizing defects on PCBs. The proposed model integrates Coordinate Attention (CoordAtt) and a Bi-directional Feature Pyramid Network (BiFPN) into its architecture to improve detection accuracy. Additionally, it outputs annotated images and visualises the defect detection process using heatmaps to facilitate effective defect localisation. The model has been evaluated using two public benchmark datasets: DeepPCB and TDD-NET. The results of the experimental evaluations demonstrate the effectiveness of the proposed Coord-BiFPN-YOLOv8 model. It achieved on the DeepPCB dataset a mean Average Precision at an Intersection over Union (IoU) of 0.5 (mAP50) of 93.4%, and for mAP50:95 the model achieved 65.9% accuracy. Also, on TDD-NET, the proposed model achieved an impressive mAP50 of 99.1% and an mAP50:95 of 67.5%. These results highlight the model’s high precision and robustness in detecting PCB defects. Additionally, the proposed model outperforms the standard YOLOv8 in defect identification, with improvements of 1.8% in F1-score and 6.4% in mAP50:95.