An improved network with deformable residual feature fusion for multi-target detection in underground coal mines

Abstract In underground coal mines, reliable multi-target detection is crucial for ensuring safety, yet it remains challenging due to uneven lighting, complex environments, and varying target sizes. Traditional deep learning methods often struggle with small object detection, leading to missed or false detections. To address these challenges, we propose DPMSNet, an improved network incorporating deformable convolution and a mixed local channel attention mechanism. First, a feature fusion enhancement module, DPC-Block, is designed, combining deformable convolution to boost the aggregation ability of local features, thereby effectively improving the fusion effect of multi-scale targets between shallow and deep features. Second, a lightweight mixed local channel attention module is introduced, which promotes the model’s attention to different object characteristics and further improves the target positioning accuracy. Last, the SIOU loss function is used to optimize the bounding box regression task, thereby accelerating model convergence. Experimental results on self-built and public DsLMF + datasets demonstrate that DPMSNet achieves an mAP@0.5 of 98.1% and 98.0%, and a Recall of 95.6% and 94.8%, respectively. Notably, the model achieves an AP@Small of 62.0% on the self-built dataset, representing a 3.3% relative improvement over the baseline. Specifically, in the demanding helmet detection task, the proposed model achieves an mAP@0.5 of 94.2%, an mAP@0.5–0.95 of 55.6%, and a Recall of 89.6%, substantially outperforming the baseline across all primary evaluation metrics. These results highlight the effectiveness of our proposed enhancements in complex underground environments.