Abstract: With the growth of service life, tunnels will inevitably undergo aging, and as an important infrastructure for the travelling of urban residents, tunnel safety inspection is crucial. At present, the crack disease on the tunnel surface is mostly detected using the images taken by cameras. However, the cracks have a small pixel percentage in the image, and its detection process is time-consuming and labor-intensive. Hence, there is an urgent need for a method that can accurately detect the cracks in a large field-of-view range. This paper proposes a learning structure based on super-resolution generative adversarial networks, which is applicable to any segmentation network, and proposes a method for efficiently constructing training data to be applied to the proposed learning structure. The performance of the proposed method is evaluated on 1606 crack images with randomly degraded quality. The results show that the crack detection IoU and F1 scores under the proposed learning structure are 63.686% and 77.811%, respectively, and the variances are 0.9008 and 0.5015, which effectively improves the performance of crack detection and has high robustness to the input data.

Key words: concrete tunnels, crack detection, super-resolution generative adversarial network, segmentation algorithm