Multiscale recognition of objects in scenes.

Summary In this thesis, we study the possibility of recognizing objects in compressed images, without reconstructing them. The most suitable compression algorithm seems to be the one based on the extraction of multi-scale staggered contours from images. The recognition problem leads us to introduce a new tool for binary image comparison: the censored Hausdorff distance. This tool has proven to be robust and fast to compute. These two points are carefully studied. This distance is finally used to recognize and localize specific objects in large scenes. We propose three multiscale approaches to solve this problem, which take into account the fact that the desired object may be partially hidden, or that it may be seen from a different angle than its model. The algorithm we have developed is fast on a classical workstation. Its robustness has been carefully studied. Its parallelization allows us to reach real time in a reasonable operational framework.