Safety Helmet Wearing Detection Based on A Lightweight YOLOv4 Algorithm

Safety helmet wearing plays a major role in protecting the safety of workers in industry and construction, so a real-time helmet wearing detection technology is very necessary. This paper makes a lightweight improvement on the original YOLOv4 algorithm and obtains a real-time and efficient safety helmet wearing detection algorithm. The lightweight YOLOv4 uses GhostNet as the backbone network for reducing the number of parameters. The adoption of the coordinate attention mechanism makes the model focused more on the region of interest and effectively distinguishes the background and foreground. Weighted bi-directional feature pyramid network (BiFPN) is introduced to combine with PANet as the enhanced feature extraction network of improved YOLOv4 to further process multi-scale information, enhance feature fusion and improve detection accuracy. Replace the general 3 × 3 convolution with the depthwise separable convolution to further reduce the network structure parameters and improve the inference speed. The experimental data show that the mAP of the improved YOLOv4 model reaches 91.30%. Compared with the original YOLOv4, the accuracy is only reduced by about 1%, but the speed is 45% higher, and the model size is 197M lower. Compared with YOLOv4-Tiny, the accuracy is improved by about 10%.