A Lightweight Model for PCB Surface Defect Detection

Traditional deep learning models for printed circuit board (PCB) surface defect detection achieve high accuracy but incur heavy computational costs, making deployment on resource-constrained edge devices in industrial environments challenging. This study aims to develop a model that balances lightweight design with high detection performance. We propose an improved object detection model based on YOLOv10n. To reduce computational load and parameter count, Ghost Convolution modules are integrated into the backbone and neck networks to replace standard convolutions. Additionally, we propose SimAM-m, an algorithmic extension of the Simple, parameter-free attention module (SimAM), which is introduced before the detection head to enhance spatial and channel-wise discriminative features while suppressing background noise without introducing additional learnable parameters. The model was evaluated on two PCB inspection tasks: component misalignment detection and solder joint defect classification. The proposed YOLOv10n-GS-m reduced parameters from 2.3M to 2.2M and GFLOPs from 6.7 to 6.3. For component misalignment, it achieved 75.2% mAP@0.5:0.95 with zero missed detections. For solder joint defects, it attained 46.1% mAP@0.5:0.95, outperforming baseline models. Integrating Ghost Convolution and SimAM-m balances high precision and lightweight requirements, effectively reducing missed detections and enhancing feature representation for automated PCB surface defect inspection.