CG-DETR-Based Real-Time Detection of Coal and Gangue Under Complex Underground Mine Conditions

Abstract Reliable identification of coal and gangue remains difficult in underground mining scenarios. To improve detection performance under such conditions, this study presents a real time CG-DETR framework for coal and gangue recognition. In actual mine environments, insufficient illumination, airborne dust, motion blur and object occlusion often weaken the visual contrast between the two materials.. These factors reduce the reliability of feature extraction and may cause missed targets, incorrect detections and repeated predictions. Specifically, depthwise convolution (DWConv) is replaced with spatial-to-depth convolution (SPD-Conv), allowing the network to preserve shallow texture cues and local spatial information more effectively while alleviating fine-grained feature loss under low-light conditions. An HSA-DARE attention mechanism is then integrated with HGBlock to construct an HGBlock-HybridAttention-Sequential module, which enhances discriminative semantic representation during deep feature extraction. Meanwhile, an independent HSA-DARE module is embedded into the backbone to further refine high-level semantic features. The original neck structure is also redesigned using a generalized feature pyramid network (GFPN), promoting cross-scale information interaction and improving the perception of dense, small-sized and occluded targets. Finally, Shape-NWD loss is adopted for bounding-box regression, thereby improving localization precision and detection robustness. The experimental results from both comparative tests and ablation studies confirm the validity of CG-DETR. The proposed model obtains 94.9% Precision, 88.7% Recall, and 93.3% mAP0.5. Its inference speed reaches 45.7 FPS. Compared with the original RT-DETR baseline, the three accuracy indicators are improved by 4.2%, 4.7%, and 5.2%, respectively. Compared with YOLOv5l, UAN-DETR and other representative detectors.