Aiming at the engineering problems such as insufficient crack detection accuracy and limited real-time performance in the complex geological environment of underground mine tunnels, an underground mine tunnel crack-segmentation network (UMTC-net) integrating multi-scale feature perception and adaptive attention mechanism was proposed. This network can realize cross-scale feature extraction of crack images from local texture to global structure through hierarchical cascading of Swin Transformer module groups. Meanwhile, a scaling cosine attention mechanism encoded by relative positions in logarithmic space was introduced to effectively suppress the interference of abnormal pixels. In addition, a codec framework based on dynamic patch merging/expansion was constructed, which solved the problems of ambiguous boundary positioning of fine cracks and high false detection rate in complex backgrounds in traditional methods. The results show that the UMTC-net has an accuracy of 85.15%, an average intersection-union ratio of 85.78%, and an F₁ value of 83.27% in the Crack 500 dataset, and an accuracy of 87.51%, an average intersection-union ratio of 79.98%, and an F₁ value of 86.95% in the MineTunnelCrack-2000 dataset. It exhibits stronger robustness, especially in low light and high dust environments. This network achieves an inference speed of 38.9 ms on the RTX 3060 mobile graphics card, occupying only 5 230 MB of memory and reducing deployment costs by more than 40%. It meets the real-time and low-power requirements of portable detection devices, and has a higher cost-effectiveness for adaptation. In the field test, the detection efficiency of UMTC-net is 8 times higher than that of manual inspection, and the missed detection rate is reduced from 18% to 3.2%. The research results provide an efficient and accurate new scheme for crack detection in underground mine tunnels, which is helpful to find potential safety hazards in time and ensure the safety of mine production and stable operation of equipment.

Underground mine tunnel  /  Crack detection  /  Multi-scale feature fusion  /  Adaptive attention mechanism  /  Mine image analysis