Abstract

The increasing complexity and miniaturization of electronic devices have heightened the need for high-quality PCB inspection to ensure product reliability and functionality. This study pro-poses a deep learning-based detection system supported by a custom-designed optical setup us-ing coaxial illumination to deliver high-contrast images with clearly highlighted defects. The data were collected from faulty PCBs during production and enriched via various augmentation tech-niques, including the creation of artificial defects, to enhance the diversity of the dataset. The en-hanced YOLOv8 framework was evaluated with multiple backbone configurations—YOLOv8+1P,ResNet152, HRNet, HRDnet+2 backbone, and HRDnet+3 backbone—achieving test accuracies of 92, 88, 90, 93.5, and 93.8%, respectively. The results indicate that HRNet and the HRDnet+3 back-bone significantly improve the detection of small and blurry defects, with HRDnet+3 providing enhanced deployment efficiency for real-time industrial inspection systems.