The French railway network has an infrastructure which is the issue of periodic monitoring to detect the presence of damages. Nowadays, this inspection is largely made visually by monitoring operators. Several companies test new vectors like UAVs (Unmanned Aerial Vehicles) in order to monitor concrete structures by acquiring photographs. These photographs, taken under constraints, can be treated in real time to identify all kinds of damages(like cracks).

This thesis focuses on the detection of these damages automatically. The main goal is to develop a system which is able to detect, localize and record potential damages at the time of acquisition. A big challenge is to identify sub-pixel damages like cracks in real time in order to improve the acquisition. To this end, a fast local analysis by thresholding was designed to process large images. To detect some types of damages, convolutional neural networks are used but the lack of data requires the constitution of our own dataset. The works on the extraction of information on images can be useful in other applications such as word spotting or intelligent vehicles.