DLE-YOLO: A Helmet Detection Model Based on Detail Capture and Precise Localization

In response to the challenge of low helmet detection accuracy in actual construction site scenarios due to factors such as worker occlusion and background interference, this paper proposes a helmet detection model based on YOLOv8, named DLE-YOLO. In the DLE-YOLO model, we introduce dilated convolution to improve the pooling operation in the SPPF structure, implement parameter sharing between convolutional layers, and propose a new DSSPPF module to better capture image details and global contextual information. At the end of the neck part, we incorporate the LSK attention mechanism to focus on regions of interest in the image, enhancing feature extraction capabilities. Additionally, we use the EIoU bounding box regression loss function to more effectively learn and adjust the shape and position of bounding boxes. Experimental results show that the proposed detection model achieves a 1.4% improvement in [email protected] and a 0.7% improvement in [email protected]:0.95 on the SHD dataset; and a 0.8% improvement in [email protected] and a 0.3% improvement in [email protected]:0.95 on the SHWD dataset. The performance of this detection model surpasses other mainstream algorithms, enabling accurate helmet detection across different complex scenarios.