Segmentation algorithms, Markovian fields and parallelism.

Summary This thesis deals with the parallelization of image segmentation algorithms, based on Markovian models. After a brief review of Markovian field techniques, we present transputer networks and more particularly the t-node, as well as its distributed operating system helios. With the help of some experiments, we show that we can predict the execution time of some algorithms on t-node machines. We then present two algorithms: the first one is a segmentation of textured images, the second one is an augmentation of multispectral images, each followed by its parallel version implemented on the t-node. The last part discusses in detail a color image segmentation algorithm, intrinsically parallel. The parallelisation possibilities of such an algorithm are discussed, but not implemented for material reasons, the t-node not being a suitable machine. This algorithm builds locally, on overlapping images, segmentations that are made to cooperate in order to obtain a global segmentation. The global coherence is done by processing a graph of labels.