A significant aspect of a subway tunnel condition assessment is the systematic inspection of the inner concrete lining state. Fractures and cracks are one of the most common problems of the tunnels in service. It is therefore necessary to detect and classify them according to their characteristics and danger level. Underground infrastructure such as metro tunnels comprise many kilometres in length. Innovative technology is vital for the efficient maintenance through detection of possible failures due to fracturing. We will present a new semi-automatic system of detection of crack presence under development for the Athens metro. The system is composed of an acquisition system (camera) geolocated, a module of cracks detection and an analysis module. In this paper, we will focus on the system core, i.e. segmentation and extraction of the fractures. The segmentation module is based on a class of artificial neural network deemed more suitable for the analysis of visual images, i.e. an efficient convolutional neural network (CNN). So the first step of this study is to find the most suitable CNN and its optimized parameters in terms of precision. Several CNN will be compared. The extraction module will first extract the segmented cracks and then define a set of parameters such as length, width, area and form. This extraction takes into account that the camera has limitations and a crack could be present on several images. Future applications of this system include the automatic surveillance of the fracture evolution in time and the application of adapted prevention measures.