Improved DN-DETR for Safety helmet wearing Detection
Yuxuan Yan, Kaijie Niu · 2023
Safety helmet wearing detection is a hot research topic in industrial production and other fields. Many models based on YOLO have made significant improvements to this task in the past. However, transformer-based DETR models are rare, and the DETR model suffers from slow convergence speed and low accuracy in detecting small objects. Therefore, a new object detection model, CDN-DETR, which combines Res2Net with DN-DETR, is proposed. The improved Res2Net is used to replace the backbone network of the DN-DETR model. To address the issue of poor training results for the attention detection module, Smooth-L1 and GIOU loss functions are linearly fused, enabling the model to converge to higher precision. To address the lack of negative sample noise training in DN-DETR, a positive-negative sample contrastive denoising training algorithm is used. Experimental results show that the improved CDN-DETR model can achieve better performance with the same number of parameters. Compared to the unimproved DN-DETR, the proposed CDN-DETR has improved the recognition accuracy by 1.8% on the Safety helmet wearing detect dataset.