Abstract: In intelligent identification of lining cracks of metro tunnels, it is difficult to provide physical information of the crack morphology through target recognition algorithms due to complex tunnel environment and diverse crack shapes. Moreover, traditional semantic segmentation algorithms suffer from mask fragmentation, large segmentation errors, and low efficiency. To address these challenges, an improved YOLOv11 tunnel lining crack segmentation model—SGE-YOLO (Spatial Group-wise Enhance-YOLO)—is proposed. The model aims to overcome the limitations of existing methods and improve detection efficiency while ensuring high recognition accuracy. The SGE-YOLO model structure adopts an optimized spatial group-wise attention mechanism to achieve the fusion and enhancement of multi-scale features at the shallow, medium, and deep layers, highlighting the key features for crack recognition and reducing interfering features. Additionally, an improved sliding inference method for tunnel lining cracks based on YOLOv11 is proposed to enhance the efficiency of crack recognition. To effectively solve the problem of mask fragmentation in crack segmentation, a crack morphology optimization algorithm based on single connected domain recognition, morphological closing operation, and skeleton extraction methods is proposed to ensure the continuity of the crack mask. Comparative experiments with various segmentation models yield the following results: (1) The SGE-YOLO model exhibits a 324% increase in crack segmentation efficiency compared to conventional YOLOv11 model, and a 75% reduction in misjudgment rate. (2) The parameters of the SGE-YOLO model only increase by 0.2 MB compared to the original YOLOv11 model, and the training time only increases by 8.4 s, which can meet the computational and time efficiency requirements in engineering. (3) After using the sliding inference and morphology correction algorithms, the misjudgment rate of SGE-YOLO is significantly reduced, the mask fragmentation is improved, and the speed of recognition and segmentation is increased by approximately 434%, proving the positive effect of the sliding inference and morphology correction algorithms on the accuracy and misjudgment rate of crack segmentation.

Key words: metro tunnel, lining cracks, semantic segmentation, multi-scale feature fusion, YOLOv11