RDC-YOLOv5: Improved Safety Helmet Detection in Adverse Weather

Outdoor construction sites are frequently affected by fog and various adverse weather conditions, resulting in a decline in the quality of the captured images. This deterioration ultimately leads to a significant drop in the performance of helmet-wear detection systems at construction sites. To address this challenge, we proposed an improved YOLOv5 model. Firstly, we set up a restoration network to effectively restore hazy image quality and enhance intricate details. Secondly, we introduced a micro-scale detection layer, enabling more efficient capture of smaller objects and effectively mitigating the adverse weather-induced variations in object scale. Finally, we added a cross-layer connection to meticulously amplify the fine-grained features of objects within the network’s shallower layers. Our comprehensive evaluation of the enhanced YOLOv5 model, conducted on two distinct datasets, yielded a final mean average precision (mAP) value of 95.1%. This substantial improvement effectively reduces both false detections and missed detections, enhances overall robustness, and significantly improves detection performance in challenging adverse weather conditions.