Résumé
Our problem is a diagnostic task. Due to environment degraded conditions, direct measurements are not possible. Due to the rapidity of the machine, human intervention is not possible in case of position fault. So, an oriented vision solution is proposed. The problem must be solved for high velocity industrial tooling machines. Degraded conditions: vibrations, water and chip of metal projections, dazzling..., are present all the time. Image analysis in constraint environment depends on constraint importance. Before tooling, the vision system has to answer: is it the right piece at the right place? Complementary methods presented in this paper are proposed in an adaptive way to solve this diagnostic problem. This detection is made by comparing an acquired camera image of the piece to be tooled with an image of reference. Some image processing methods are performed and combined in order to extract 2D features. Some of these 2D features are evaluated and used as parameters in a diagnostic process. After a data analysis, image parameters are reduced. First, we present an overview of image processing methods generally used to solve that kind of problem and their limitations in our particular degraded case. In order to obtain automatic and robust classification, two methods are implemented. The first one is based on Bayes technique that provides a good classification in case of fault presence. The second method is based on neural networks and provides good results in case of images without faults. These two methods give a global rate of good classification greater than 90\%, for 720 images acquired from an industrial site.