YOLO-HCS: a YOLO-based framework for small defect detection in industrial automated optical inspection

Abstract Connectors are used in electronic and industrial systems to establish electrical connections between components, with terminal surfaces playing a role in ensuring current transmission. Accurate defect detection on these terminals is vital for maintaining product reliability and quality. However, detecting small, irregular, and scale-varying defects remains a challenge for automated visual inspection systems. To tackle this challenge, this work presents You Only Look Once with Hierarchical Coordinate Sampling (YOLO-HCS), a defect detection framework that integrates hierarchical positive sample selection (HPSS) to improve training stability, coordinate attention (CA) to enhance spatial localization of small targets, and scylla-Intersection over Union (SIoU) loss to refine bounding box regression through geometric constraints. We validate the approach on two datasets collected using automated optical inspection (AOI) systems. The proposed YOLO-HCS framework integrates hierarchical positive sample selection, coordinate attention, and SIoU loss to improve the detection of small and irregular defects in industrial AOI environments. Experimental results on two real-world AOI datasets demonstrate that the YOLO- HCS model achieves higher mean average precision (mAP), outperforming YOLOv7 and other state-of-the-art detectors. Beyond accuracy gains, the proposed method improves production line reliability, reduces reliance on manual inspection, and satisfies the real-time speed requirements of high-throughput manufacturing. Its lightweight design enables efficient deployment in industrial AOI systems while maintaining real-time inspection capability. While validated on connector terminals, the approach is broadly transferable to semiconductor packaging, printed circuit board (PCB) inspection, and precision component manufacturing, offering a practical pathway to strengthen industrial quality control and reduce operational costs.