Deep learning based distortion restoration and defect segmentation from linear scanning camera images

Abstract With long-term operation of tunnels, the surface of tunnel lining is prone to crack, spalling, water leakage and other defects. Therefore, the detection of tunnel surface defects has become an important process to ensure the safety of tunnel operation. Linear scan cameras have become the main way of tunnel surface defect detection because of their ultrahigh resolution and ultrahigh scanning frequency. However, in the process of collecting tunnel surface images at high speed, the linear scan cameras are easy to cause images distortion, which leads to reduction of tunnel defect identification accuracy. Thus, this paper proposes a tunnel surface distortion image restoration method based on a supervised generative adversarial network, which introduces an attention mechanism to guide the method to calibrate the weights of feature channels. It solves the problem of distortion in the process of images acquisition of the linear scan cameras. Then, to deal with the difficulty in identifying small-size tunnel defects in complex tunnel environments, this paper combines a backbone network with a multi-scale feature fusion module based on the You Only Look At CoefficienTs (YOLACT) network, which enhances the feature extraction ability of the network and improves the accuracy of tunnel surface defect identification.