Defect Detection in Printed Circuit Boards Using YOLO Optimized by Ghost Convolution

Abstract

Automated optical inspection (AOI) is essential for ensuring the reliability of printed circuit boards (PCBs), particularly as manufacturing processes evolve toward higher component density and structural complexity. This work presents a systematic, architecture-dependent, and hardware-aware evaluation of Ghost Convolution applied to YOLOv10, YOLOv11, and the attention-centric YOLOv12 for PCB defect detection. Unlike previous studies that typically design a customized lightweight detector around a single architecture, a consistent Ghost-based substitution strategy is applied across three distinct backbone designs to investigate how architectural characteristics influence detection performance and computational efficiency. Experiments were conducted on the TDD-Net and DsPCBSD+ datasets using the nano, small, and medium variants. The evaluation considers precision, recall, mAP@50, inference speed, parameter count, and GFLOPs. Convergence behavior was examined by comparing the epochs associated with the maximum validation mAP@50, while representative YOLOv11m predictions on DsPCBSD+ were qualitatively analyzed to identify successful detections and difficult cases. The small variants were additionally deployed on an NVIDIA Jetson Orin Nano 8 GB and evaluated in terms of latency, FPS, global throughput, process RAM, and power consumption. The results show that YOLOv12 provides the highest detection accuracy on TDD-Net, whereas YOLOv11 offers greater robustness on the more challenging DsPCBSD+ dataset. Ghost Convolution does not produce a uniform convergence effect, with its influence varying across architectures and model scales. The qualitative results confirm successful localization across different defect sizes and geometries while revealing remaining difficulties involving subtle, partially visible, and visually similar defects. YOLOv10 translates Ghost-based substitutions more effectively into workstation inference-speed gains, whereas YOLOv12s+Ghost provides the most balanced embedded improvement by enhancing latency and throughput while reducing power consumption. Overall, the findings demonstrate that the effectiveness of Ghost Convolution depends on both the backbone design and the target deployment environment, highlighting the importance of architecture-aware lightweight optimization strategies for real-time PCB inspection systems.