Small target detection algorithm for underground helmet based on REIW-YOLOv10n

To address the issue of low accuracy in detecting small targets, such as safety helmets, in complex underground working environments, caused by various factors like lighting conditions and equipment obstructions, a new image detection algorithm for small safety helmet targets based on the REIW-YOLOv10n model is proposed within the framework of the YOLOv10n model. The REIW-YOLOv10n model consists of four parts: Input, Backbone, Neck, and Head. To enhance the model’s ability to extract multi-scale features, the RepNMSC structure is designed, and the C2f structure in the Backbone section is improved, which boosts the model’s capability to extract features of multi-scale safety helmet targets. To preserve the small target semantic information in the Neck section, the ERepGFPN structure is adopted. This structure uses cross-layer connections to process high-level semantic information and low-level spatial information with the same priority, achieving the integration of small target features. Then, in the Head section, a P2 small target detection head is added, and the P5 large target detection head is removed. This improves the model’s detection performance for small safety helmet targets in underground environments while maintaining a lightweight model. Finally, the MPDIoU loss function is optimized using the concepts of Inner-IoU and Wise-IoU v3. Inner-Wise-MPDIoU employs scaling factors and gradient gain strategies to accelerate model convergence. Experiments using the CUMT-HelmeT dataset demonstrate that compared to YOLOv10n, REIW-YOLOv10n improves mAP@0.5 by 5.73%, reaching 88.24%. Compared with other mainstream YOLO series algorithms, such as YOLOv5s, YOLOv7-tiny, YOLOv8n, and YOLOv9-tiny, REIW-YOLOv10n outperforms them in terms of accuracy and model weight size, offering the best overall detection performance. REIW-YOLOv10n significantly improves the accuracy of detecting small safety helmet targets in complex underground environments while balancing lightweight design and real-time processing, making it convenient for deployment on underground edge devices.