Improved Safety Helmet Detection Model Based on RTDETR: RTDETR-IHD

This paper proposes a method primarily based on the latest deep learning technologies to enhance the performance of the RTDETR model for specific detection tasks by improving its key components. Firstly, we improve the model’s backbone structure using the inverted Residual Block (iRMB) technique, which effectively enhances feature extraction efficiency and accuracy by altering the connection mode of the residual blocks. Secondly, we introduce an efficient attention mechanism, replacing the multi-head self-attention module in the model, to increase the model’s ability to focus on critical features, thereby improving detection accuracy. Finally, we employ a dynamic upsampling method to optimize the feature map restoration process, dynamically adjusting the upsampling strategy to effectively enhance the model’s capability in detecting small objects. Experimental results on the safety helmet detection dataset demonstrate that the improved model significantly boosts detection accuracy and efficiency while maintaining real-time performance. Compared to the original RTDETR model, the enhanced model achieves substantial improvements in mAP, Precision, and Recall values, while reducing the total computational cost from the original 57.3 GFLOPs to 49.6 GFLOPs.