Research on mine safety helmet detection algorithm based on multi-module collaborative optimization

To address the high false-negative rates, poor adaptability to complex environments, and deployment challenges associated with traditional manual inspections and existing YOLOv8 models in mine safety helmet detection, this study implements targeted improvements to the YOLOv8 algorithm and constructs an adapted model tailored to the specific characteristics of mine environments—such as low illumination and the prevalence of small-object helmets. Methodologically, the C2f-FE lightweight module replaces the original C2f module. It combines FasterNet Block with EMA attention mechanisms to balance small-object feature extraction and model lightweighting. The Dy-RepGFPN feature fusion network is introduced, utilizing dynamic upsampling and CSP_fusion modules to aggregate multi-scale features and suppress environmental interference. We innovated the UCDN-Head detection head, leveraging parameter sharing and independent BN layer calibration to optimize detection accuracy and environmental adaptability. Experiments conducted on a self-built dataset of 5,420 mining safety helmets yielded the following results: the improved model achieved mAP@0.5 of 84.4%, an increase of 5.0% compared to the YOLOv8 baseline model; mAP@0.5-0.95 was 81.5%, precision was 88.5%, and recall was 89.8%, with 3.1M parameters, 6.1 GFLOPs, and 80.5 FPS. The conclusions demonstrate that the improved model achieves comprehensive advantages of “high accuracy, lightweight, and high adaptability” for mine safety helmet detection, meeting real-time monitoring requirements in mines and providing technical support for head protection monitoring of underground personnel.