TR-YOLO: An Enhanced Model for PCB Mixed Defect Detection and Classification

Accurate detection and classification of mixed defects in printed circuit board (PCB) are critical for maintaining the quality and reliability of electronic product. This paper presents TR-YOLO, an enhanced version of the YOLOv5 model, which is developed for PCB defect inspection. TR-YOLO introduces a modified backbone structure incorporating a C3TR module, enabling effective extracting features at multiple scales while minimizing computational resources overhead. The findings from the experiments indicate that TR-YOLO attains a mean average precision (mAP) of 96.4%, with 4.1 giga FLOPs and 1.80 million parameters. Compared to the YOLOv5 model, the proposed TR-YOLO improves mAP by 0.63%, precision by 0.73%, recall by 1.04%, F1-score by 1.34%, and overall accuracy by 0.17%. Moreover, it reduces computational complexity by 2.38% (i.e., reduced FLOPs) and enhances the speed by 7.61%. These findings highlight TR-YOLO’s ability to deliver a balanced performance between detection accuracy and processing efficiency. The proposed model shows strong potential for real-time automatic optical inspection (AOI) in PCB manufacturing, contributing to improved product quality and production reliability.