Helmet Detection in Mines Using Two-Branch YOLOv5 Network with Adaptive Weight Adjustment

Helmet detection in mines is a significant challenge due to poor visibility and complex environments. However, the head-shoulder feature of miners, which is more stable and easier to identify than other body parts, can supplement head detection to improve helmet detection accuracy. In this paper, we first use SRGAN to reconstruct mine monitoring images with super resolution to obtain more abundant details. Next, we investigate the head-shoulder and head characteristics of miners, utilizing a two-branch network structure to regress two bounding boxes simultaneously. YOLOv5 is employed to predict the head-shoulder and head regions for each independent subnet. Additionally, unlike traditional methods that set fixed weights for different features, we design an adaptive weight adjustment mechanism that dynamically adjusts the relationship between head-shoulder and head predictions, resulting in more accurate helmet positioning. We use real underground mine surveillance images from the CUMT-HelmeT dataset to verify the effectiveness of our proposed method. Experimental results demonstrate that our two-branch detection method achieves excellent performance in detecting safety helmets in mines, particularly under low lighting conditions.