An intelligent coal gangue recognition method based on improved YOLOv12

To address the difficulty of accurately and efficiently recognizing coal gangue caused by complex environmental factors such as high dust concentration and highly variable illumination in mines, this study improved the YOLOv12 network model and proposed an intelligent coal gangue recognition method based on improved YOLOv12. A Dual-Scale Sparse Attention (DSSA) mechanism was designed to enhance the model's attention to multi-scale coal gangue target regions and its spatial perception capability. A Multi-Condition Feature Refinement (MCFR) mechanism was designed to perform condition-guided fusion of deep and shallow features, which effectively enhanced the discriminative representation between coal and coal gangue. A Dynamic Multi-Task Balance Loss (DMTBL) function was constructed to achieve adaptive weight adjustment among localization, classification, and confidence, thereby strengthening the model's learning capability for hard sample regions. Experimental results showed that the improved YOLOv12 achieved a precision, recall, and mAP of 96.5%, 94.9%, and 95.8%, respectively, in the coal gangue recognition task, representing improvements of 3.8%, 4.5%, and 4.5% over the original YOLOv12, which effectively addressed issues such as missed detection, false positives, and blurred boundaries while maintaining a high inference speed of 47.7 frames per second. Visualization results of activation heatmaps showed that the improved YOLOv12 accurately focused on the target object regions when processing coal gangue with different structures and texture complexities, with no obvious background interference, and the activated regions basically cover the main contours of coal blocks and coal gangue.