Abstract: Intelligently identifying cracks in metro tunnels is challenging due to factors such as short skylight detection time, low manual detection efficiency, inconspicuous crack characteristics, and various disturbances in the tunnel. Therefore, a modified YOLO (M-YOLO) algorithm based on YOLOv8 is proposed to efficiently and intelligently identify cracks in tunnels. This algorithm adopts full-dimensional dynamic convolution to replace the traditional convolution module, remarkably improving the detection accuracy and avoiding model parameter expansion. The C2fGC module is introduced to improve the network structure, establishing a new feature extraction and dimensionality reduction mechanism and enhancing the high-level feature representation. The convolutional block attention module attention mechanism module is introduced to strengthen feature learning and extraction of crack region images and reduce background interference, effectively improving the detection accuracy. The WIOU loss function is introduced to adjust the degree of penalty of geometric factors, which improves the generalization ability of the model and helps it perform better even with low-quality data samples. The test results show that in the real samples, which are used for identifying cracks in a metro tunnel, the mean accuracy of the M-YOLO algorithm is as high as 83.0%, which indicates a significant improvement of 15.7% compared with the original model.

Key words: metro tunnel, crack identification, object detection, dynamic convolution, attention mechanism