Deep Learning for automatic crack detection inside tunnels - Second Prototype

The rapid urbanization of cities around the globe and the corresponding exponential growth in transportation infrastructure to support ever increasing population densities, has led to the corresponding increase in road tunnels as a means of alleviating congestion, where and when possible. This necessitates development of automated techniques for detection of cracks on the surfaces inside tunnels given the high costs of manual visual inspection both monetarily and in terms of time. Structural integrity testing of the road tunnels imposes the first problem of managing the traffic flow during the inspections, since it may not always be possible to block the traffic and examine the entire tunnel to perform a human visual inspection or to prepare and mount a complex robotic structure for the same purpose. In this scenario, a low-cost automatic system can be a smart solution and can overcome many of the previously described problems. This thesis project is centered in the University Transportation Center, a collaboration between different American universities (California State University Los Angeles, Colorado School of Mines and Leigh University) focused on solving challenging problems in the urban infrastructures, such as galleries, buildings or bridges and financed by the Department of Transportation of United States. In particular, this work introduces a low cost automated system for tunnel inspections, which is also translated in a very simple structure that does not require so complex preliminary work and can be used also in urban tunnels with normal traffic flow. This means that this approach is completely different from the ones present in literature, specifically in terms of cost. The system is composed by two main blocks: the acquisition block, for the collection of images in the galleries surfaces, and the decisional block, composed by a deep convolutional neural network which has to classify properly the images in two main categories, Crack Detection and No Crack Detection. In the acquisition system, a second generation prototype is based on a picture acquisition system completely automatized and able to collect a huge amount of pictures autonomously without any kind of post processing. In the software side, thanks to deep learning techniques, it has been possible to exploit the power of Inception-v4, a deep network built by Google which can be retrained to respond properly to the specific purpose of the crack detection. Moreover, a second deep network has been built from the scratch and trained on the same purpose, in order to make comparisons between the two approaches and in order to choose the most suitable one. Both the networks have been trained with a dataset obtained through the acquisition system and by testing different parameters tuning, so that an optimal solution can be found. All the results are then shown at the end of the dissertation. The thesis is organized with a first theoretical introduction on the main aspects of Deep Learning and Convolutional Neural Networks, in order to have a strong background to properly understand the work behind the project. The crack detector is then presented, with a first focus on the acquisition system and then on the networks. The test results are presented and a final conclusion estimates the different approaches and tries to define the correct settings of the system, in order to have a prototype ready to be used on real time inspections in an actual urban tunnel.