Segmentation adaptative pour le codage d'images

This Ph.D. dissertation analyses the problem of segmenting an image into a set of regions corresponding as much as possible to the real objects of a scene. The goal of this segmentation is to go from a numerical representation of an image to a symbolic one, i.e. the regions and their characteristic features. Starting from the set of picture elements, one reaches a more compact model where region frontiers define the contours of the objects and the signal within each region their texture. Such a representation can find useful applications for scene understanding and image coding. Recent works have shown the potential of a contour-texture model for image coding due to the properties of our human visual system. We have studied here the way of coding a segmented representation of an image when using high order polynomials for approximating the different regions. The proposed segmentation algorithm is adaptive. Given a certain approximation to model the texture within each region, one tries to modify adaptively the region shape and the approximation parameters. To do so, the image is first split into a set of squares of different sires in order to obtain an optimal correspondence between the original signal and its approximation within each square. Then starting from this initial partition, adjacent regions are iteratively merged till one reaches a segmentation with a certain number of regions of any shape. At each step of the merging process, the two most similar regions are associated on the basis of the adequacy of the approximation on the new region. Polynomials of degree 0 to 3 have been used in the approximation process. Once the final segmentation is obtained, frontier information and texture information are coded separately. Performances for redundancy reduction are impressive: acceptable quality images can be obtained with compression ratios of the order of 30 to 1. It is shown how most of the semantics can be preserved with coded pictures at compression ratios ranging from 60 to 1 to 130 to 1. The algorithm has been applied to three different 256x256 natural images quantized with eight bit dynamics.