Research on Improved YOLOv8 Based on EMA and DyHead for Mine Target Detection

Abstract

As an important part of intelligent mine construction, target detection technology is crucial to ensure mine safety production. However, in practical applications, due to the drastic changes in lighting, severe dust interference, small target size and complex dynamic background in the mine environment, the information obtained by traditional detection algorithms often has large errors, leading to missed detection and false detection. In order to solve this problem, this study proposes a YOLOv8-ED model enhanced by EMA attention mechanism and DyHead dynamic detection head. By embedding EMA attention after the C2f module of the backbone network, the model's ability to extract key features in low-light and high-occlusion scenes is enhanced. The original detection head is replaced with DyHead to realize unified modeling and dynamic feature fusion in three dimensions of scale, space and task. Experiments on the self-built mine violation dataset and PASCAL VOC2012 dataset show that the mAP@0.5 of YOLOv8-ED is improved by 3.2% and 2.0% respectively compared with the original YOLOv8. This research provides an effective technical solution for intelligent mine safety monitoring and has important significance for improving the overall performance of mine detection systems.