Deep Learning with Spatial Constraint for Tunnel Crack Detection

Cracks are the most common defect on the surface of tunnels, which potentially brings threaten to the safety of the tunnel and the running vehicles. Timely repairing of the crack is of critical importance. In the past two decades, various vehicle platforms have been developed on the purpose of efficient crack detection and maintenance. With these platforms, images can be captured in a traffic speed, and automatic methods can be developed for fast crack localization. However, for image-based crack detection, traditional methods often meet difficulties in handling cracks with low contrast and poor continuity. In this paper, deep learning based techniques are exploited for feature learning and representation for crack detection. A novel deep neural network is presented for pixel-level crack recognition. Hierarchical features in different stages of the convolution are fused together to overcome the influence of noise and a spatial constraint placed on the target pixels is used to guarantee the crack continuity. In the experiment, a tunnel crack dataset is constructed for performance evaluation. Experimental results demonstrate the effectiveness of proposed method.