SF-YOLO: designed based on tiny feature for PCB surface defect detection and deployment in embedded systems

Owing to the minute nature of printed circuit board (PCB) surface defects, their precise detection poses significant challenges. To improve the accuracy of detecting small PCB defects, this paper proposes an SF-YOLO algorithm based on the improved YOLOv8 framework. A tiny detection head was designed to enhance the detection accuracy of small targets and reduce data requirements and model size. Tests demonstrated that the improved model increased mAP50 by 0.6% and mAP50–90 by 1.6% compared to the original model. The model size and parameter count are reduced by approximately 14.2 and 17.2%, respectively. The trained model was deployed to a Jetson Nano to realize a high-performance detection system for portable PCBs. The proposed method offers several advantages, including fewer parameters, a smaller model size, and ease of deployment. It is portable and satisfies the requirements of industrial applications.