A deep learning-based approach for refined crack evaluation from shield tunnel lining images

Abstract This paper develops a deep learning-based approach that extends the PANet model by adding a semantic branch which refines the process of crack evaluation to reduce inaccuracies associated with crack discontinuities and image skeletonization. The PANet model is used to segment cracks from shield tunnel lining images, and an A* algorithm incorporated into the semantic branch computes the width and the shortest length of each crack from the segmented binary images. A comparison of experimental results shows that the performance of the proposed approach is better than those of Mask R-CNN, U-Net, and DeepCrack. The proposed approach also demonstrates its superiority at mitigating crack disjoint problems and skeletonization error. The error rates of length and width quantification for the A* algorithm are lower than those for the medial-axis-skeletonizing algorithm. The evaluation metrics indicate that the proposed model is an alternative approach to segment and quantify cracks in the field.