Use of Deep Learning for Automatic Detection of Cracks in Tunnels

Tunnel cracks on concrete surfaces are one of the earliest indicators of degradation, and if not promptly treated, they could result in full closure of an entire infrastructure or even worse in a structural failure of it. Visual inspection, carried out by trained operators, is still the most commonly used technique, and according to the literature, automatic assessment systems are arguably expensive and still rely on old image processing techniques, precluding the possibility to afford a large quantity of them for a high-frequency monitoring. So, this article proposes a low cost, automatic detection system that exploits deep convolutional neural network (CNN) architectures for identifying cracks in tunnels relying only on low-resolution images. The trained model is obtained with two different methods: a custom CNN trained from scratch and a retrained 48-layer network, using supervised learning and transfer learning, respectively. Both architectures have been trained and tested with an image database acquired with the first prototype of the video acquisition system.