Automatic Detection Method for Surface Diseases of Shield Tunnel Based on Deep Learning

Abstract

In order to achieve high-precision pixel-level detection of multiple surface diseases in metro shield tunnels, a semantic segmentation model SU-ResNet++ based on deep learning is proposed. First, the encoder SE-ResNet50 based on residual unit and attention mechanism is designed and pre-trained, using as the backbone network of U-Net++ to design a new neural network model. Then, through original data collection, data preprocessing, and manual annotation, a shield tunnel surface multiple diseases dataset with 4 500 pictures is constructed. Finally, the proposed method is trained, verified, and tested on a dataset, and applied to practical engineering detection, achieving high-precision pixel-level diseases semantic segmentation. The experimental results indicate that the proposed SU-ResNet++ algorithm is applicable to the detection of shield tunnel disease data, and can automatically and accurately identify the disease category and form. Compared with the traditional semantic segmentation models, its disease identification precision is significantly improved, which meets the practical engineering requirements.

Keywords

• metro shield tunnel
• dataset construction
• semantic segmentation
• deep transfer learning
• u-net++ networks