YOLO-based Lightweight Drill Detection for Coal Mines

Abstract

Aiming at the challenges of insufficient illumination, heavy dust interference, large scale variation of targets, and limited computational resources of edge devices in underground coal mine drilling scenarios, this paper proposes a lightweight object detection model named YOLO-Drill. The proposed model is developed based on the YOLO framework with several improvements. First, a Lightweight Drill Feature Block (LDFB) is designed by integrating depthwise separable convolution and directional perception mechanisms, which reduces computational complexity while enhancing the representation ability of slender targets. Second, a Drill-Aware Feature Pyramid Network (DAFPN) is constructed to achieve efficient multi-scale feature fusion through bidirectional cross-layer interaction and attention weighting, thereby improving the detection performance of small and occluded targets. In addition, a scene-adaptive data augmentation strategy is introduced to enhance the robustness of the model under low-light, high-dust, and low-contrast conditions. Finally, the CIoU loss is adopted to optimize bounding box regression and improve localization accuracy. Experimental results on the self-constructed Drill Scene Dataset (DSD) demonstrate that the proposed YOLO-Drill achieves 85.2% mAP@0.5 and 51.0% mAP@0.5:0.95, which outperform the baseline by 6.6% and 5.8%, re-spectively. Meanwhile, the model maintains a lightweight structure with only 6.0M parameters and 19.9 GFLOPs, achieving a real-time inference speed of 87.9 FPS. The results verify that the proposed method effectively balances detection accuracy and computational efficiency, showing strong applicability in underground coal mine drilling scenarios.