Tunnel Crack Detection Method Based on Improved CenterNet

Crack disease is a major safety hazard in the daily operation of the tunnel. In serious cases, it may hinder the traffic, or even completely interrupt the whole traffic line, causing a bad impact on the social life and the environment. Therefore, it is very important to accurately detect the tunnel cracks. This paper studies the tunnel crack detection method based on CenterNet, and proposes two improvement methods for the network structure level. First, the method of multi-scale feature fusion is adopted. In the process of upsampling, the low-level semantic features obtained by downsampling are added to enhance the expression ability of the upsampling features. The second is to replace the deconvolution operation in the upsampling module with sub-pixel convolution to obtain more detailed feature map. The experimental results show that the improved CenterNet network has a significant improvement in the mAP value, thus verifying the effectiveness of this network for the tunnel crack detection, and providing a new technical means for the structural health monitoring of the tunnel.