Skip to Main Content
Describes an application of neural networks to the discovery of road surface distresses (cracks, etc.) from video sequences. After describing the context of this application, we detail its progressive design. We initially thought of using only one neural network to totally analyze each image extracted from a video sequence. We later thought of a more simple neural network analyzing only a small part of each image each time, with a preprocessor scanning the image. We finally preferred to simplify the role of the neural network by putting an image preprocessing sequence in front, to extract objects that were then identified by the neural network. We describe the sequence of treatments that we use and detail each processing step: improvement of the original image, extraction of significant objects (possible distresses) and identification of these objects by the neural network. We conclude by evaluating the performance of our system and by discussing possible improvements.