Automatic Recognition of Highway Tunnel Defects Based on an Improved U-Net Model

It is critically important for the safe operation of highway tunnels that defects in tunnel linings be promptly identified. Recently, many effective identification models based on deep learning technology have emerged in the field of tunnel defect recognition. However, the results of previous studies indicate that imperfections exist in these models in which the boundaries of the defects can only be generally located in the complex visual environment of tunnel linings, which include noisy patterns and uneven light sources. In order to more accurately identify defect details in such complex environments, an improved U-net model including a combined Squeeze-and-Excitation (SE) and ResNet block is proposed and evaluated in this paper. The accuracy of the conventional U-net model is found to be considerably improved by the inclusion of the SE and ResNet blocks. An under-sampling strategy is utilized to address the problem of imbalanced crack image data and determine an appropriate ratio of negative to positive samples for the crack dataset. The effectiveness of the proposed model is then demonstrated using actual highway tunnel lining images to compare its results with those of the Gabor Filter, Multiscale DCNN, FCN, and conventional U-net defect recognition models. The accuracy of the proposed model was found to be quite competitive with extant models.