Abstract: To enhance the accuracy of identifying surface cracks in subway tunnels, we propose an algorithm for recognizing and extracting crack features in subway shield tunnels taking into account the presence of obstructions based on deep learning and machine vision technology. We employed the convolution neural network to extract the features of cracks using image restoration technology, thereby obtaining accurate information on the cracks. First, we proposed a U-net-based method of segmenting the occlusions on the subway tunnel segment surface, marking the position information by a mask map. Then, we built a Faster R-CNN network model using VGG-16 as the feature extraction network to locate and mark the cracks in the tunnel segment image. By combining the results of the two, we can determine the relationship between the cracks and occlusions. Finally, we proposed the deep learning convolution neural network model in which the occluded crack information can be restored. The trained network model can predict the cracks in the occluded area. The experimental results demonstrate that the accuracy rate of the proposed method reaches 90% in repairing the occluded crack information, indicating its application value in actual tunnel disaster detection and comprehensive evaluation.

Key words: shield tunnel, surface crack identification, feature extraction, deep learning, machine vision technology, image restoration