Improved YOLOv5s-CBCG Algorithm for Detecting Safety Helmets in Underground Coal Mines

In the mining industry, underground workers are mandated to wear safety helmets due to numerous potential risks. However, due to that the complex and harsh environment of underground coal mines, the object images obtained by video surveillance systems suffer from challenges such as uneven lighting, small and easily obscured targets, and significant environmental interference. Consequently, most existing deep learning-based object detection methods encounter substantial difficulties in accurately, efficiently and timely detecting safety helmets in underground coal mine. These challenges include low detection accuracy, poor model robustness, and a contradiction between the enhancement of detection performance and increased computational consumption demands. Therefore, to address these limitations, this work investigates a new detection algorithm for coal mine safety helmet wearing based on YOLOv5s, named YOLOv5s-CBCG. Firstly, an enhanced feature extraction network named FEN-CA is developed by incorporating the coordinate attention mechanism, which contributes to gain more powerful feature extraction ability for small object and suppress the interference of background noise in small target images. Secondly, it designs a new feature fusion network named FFE-BCG to enhance the multi-scale feature fusion, and further improve the small object detection accuracy while reducing computational cost. Then, the strengthened feature extraction network and feature fusion network are used as the backbone and neck of YOLOv5s-CBCG, respectively. Finally, self-built M-Helmet dataset is used to conduct extensive experiments, and the results indicate that the YOLOv5s-CBCG can reach 95.90% mAP on M-Helmet dataset with less computational expense than YOLOv5s, which outperforms other comparative methods. Specifically, the mAP is 4% and 7.9% higher compared to the latest YOLOv11n and YOLOv12n algorithm.