Lightweight structure-aware detection of coal and gangue under complex operating conditions

Abstract

Accurate coal-gangue detection under complex operating conditions is crucial for reliable online perception in underground coal handling. To address degraded imaging, dense stacking, adhesion, and occlusion, this study proposes a lightweight structure-aware detector, LCGA-YOLO11, based on YOLO11. Without changing the end-to-end detection paradigm or multi-scale output mechanism, PP-LCNet is introduced for lightweight feature extraction, CoalGangueAM is used to enhance structural and boundary representation, and imWIoU is adopted to improve localization stability. The dataset was constructed from Sima Coal Mine of Lu’an Group and comprised 7500 coal-gangue images after augmentation, including 6000 images for training and 1500 images for validation. Experimental results show that LCGA-YOLO11 achieves an mAP@0.5 of 95.94%, with Precision and Recall of 93.46% and 90.36%, respectively. Meanwhile, the number of parameters is reduced from 2.59 to 2.33 M, and GFLOPs decrease from 6.44 to 5.28. The proposed model achieves a better balance between detection accuracy and computational efficiency, providing a reliable method for online coal-gangue detection under degraded underground imaging conditions.