Markov fields and contour matching in imaging.

Summary This thesis has for central theme the pattern recognition in image analysis. It is composed of three independent parts. A contour detection algorithm developed with olivier catoni is presented in the first part. We build by smoothing and thresholding binary images whose edges contain the contour points. We regularize them with a gibbs sampler, which gives us smooth contours. The algorithm operates at different scales, which we combine to obtain a multi-scale detector. The problem that we deal with in the second part is to recognize and position on a digital image given manufactured objects. The objects present on the image are the transforms of objects modeled by any similarity. It is thus a question of associating each object present on the image with the good model and of estimating this similarity. Each object is represented by a value graph. Any geometrical correspondence between two objects is translated into a correspondence between graphs that must satisfy some local constraints. We place ourselves in a Markovian context to determine a graph matching that best satisfies these constraints. The third part presents a study carried out with alain trouve on the parallelization of simulated annealing for spin glass energies. Each spin is linked to a processor. At each step, each processor is active with a probability p and all active processors renew their spin values simultaneously. We show that if all the processors are active, there is no convergence towards the global minima of the energy. As soon as p is strictly less than 1, there seems to be convergence towards energy configurations close to the minimum energy. Finally, we compute an effective time saving of the partially parallel annealing compared to the sequential annealing.