DM-YOLOv5:An Improved Safety Helmet Detection Algorithm at Complex Construction Scenes

At complex construction sites, detecting whether workers are wearing safety helmets is often affected by occlusion and sunlight change. In addition, the helmets are relatively small, which can easily lead to missed detection and false detection. This paper proposes an improved YOLOv5 model. Firstly, we added a 152×152 detection layer for scale of small objects, enhancing its feature extraction capabilities. Secondly, we added the Multi-Spectral Channel Attention (MSCA) mechanism to enhance feature learning for small objects, addressing the information loss problem caused by Global Average Pooling (GAP). Finally, different activation functions, bounding box regression loss functions, and post-processing methods were adopted for YOLOv5. The Mish function serves as the activation function to enhance the model’s generalization ability and mitigate the gradient vanishing problem associated with the Leaky Relu function. The CIoU loss was employed as the loss function to improve bounding box regression accuracy and accelerate prediction convergence. DIoU-NMS, instead of NMS, was used as the post-processing method to enhance the filtering ability for prediction boxes of occluded small objects. We tested our algorithm on a dataset with 7000 images and achieved an mAP of 95.1%, surpassing the YOLOv5 model by 2.9%. Our method significantly improves safety helmet detection performance at complex construction scenes.