Multiscale Markovian fields: applications to textured image segmentation and multi-film fusion.

Summary This thesis is composed of two independent parts. The first part deals with texture representation and segmentation of textured images. We use Gaussian modeling and Gabor filters to characterize the textures. By our analysis, we see that, from a stochastic point of view, the characterization by gabor filters and the one by gaussian modelizations are comparable in the sense of random fields, although their motivations are very different. We present a method for the segmentation of multiscale unsupervised textured images, using partitions by related regions. We use Markov field methods to segment images, which by relaxation, tend to group points with similar texture characteristics. Our method progressively limits the domain of exploration of segmentations in the natural progression of the multiscale algorithm. The second part of this thesis deals with the fusion of a series of images (x-rays) obtained from distinct films in order to restore the 2d image of an object. This requires the estimation of the transfer function allowing to pass from one film to another. This task is made difficult by the presence of spatial inhomogeneity. We propose a method to estimate the transfer function between two films of different sensitivities and non-homogeneity. We consider that the image of the object is a realization of a Markov field whose a priori energy is built on the fact that the image verifies some regularization constraints. The restoration is then done by likelihood maximization.