Abstract: Due to the fact that most of the cracks do not have distinctive features and are affected by cable,scratches,cobwebs and other linear interferences inside the tunnels,the detection effect of the existing crack detection methods,using high-resolution images,still needs to be improved. This paper takes the lining cracks as the research object,realizes the non-destructive data acquisition of the tunnel surface information based on the tunnel camera system,and acquires the high-resolution image data of 4096×2168 pixels. And we clarify the interference factors of crack identification,and constructs the interference data set and the real texture data set based on the characteristics of the disease; takes the Mask R-CNN model as the baseline framework,and adopts the K-means and genetic algorithm to optimize the parameters of RPN network. The detection effect and performance of this paper's algorithm are illustrated using comparative and ablation experiments. The results show that the algorithm proposed in this paper can realize the recognition and length measurement of tunnel cracks under high-resolution images,with lower probability of leakage and false detection,and has better detection performance for the slender and less obvious cracks,and the measured values of the cracks can provide reference information for the operation and maintenance of the subway.

Key words: Mask R-CNN, high-resolution images, size clustering, crack, length measurement