The development of applications such as picturephone, multimedia or videoconferencing has created the necessity of coding image sequences at very low bit-rates. Object-oriented image compression systems, also called second-generation methods, offer a promising alternative to classical methods. These methods are close to the behavior of the human eye, to which coding systems are finally destined, and describe an image in terms of the objects present in the scene. Thus, an object-oriented coding system operates in two steps: first, a step of analysis, called segmentation, which splits up the image into objects (sets of pixels satisfying a condition of homogeneity) and second a step modeling the shape and content of the segmented objects.
The work of this thesis is focused on the segmentation of image sequences, knowing that the final objective is to integrate it into the object-oriented coding system.
At the start, our work was based on the segmentation paradigm of mathematical morphology: the watershed of the gradient from the markers of the regions to be extracted. This scheme has resulted in numerous applications but, in the field of coding, it leads to real difficulties. These difficulties are due to the presence of fine details, or to the presence of fast motion that generates fine objects along the time axis, which amounts to the same. Indeed, in the field of coding, the conditions of morphological sampling are not satisfied. The analysis of the difficulties encountered has led us to another segmentation paradigm now based on region merging.
Thus, we have developed a region merging algorithm. Implemented onto a graph structure whose nodes and edges are valuated, the algorithm is both efficient and flexible enough to take a large variety of criteria into account. In particular, in order to make the whole coding system consistent, we have considered among merging criteria the capacities of coding algorithms applied after segmentation.
Finally, we have illustrated the potential offered to a coding system by segmentation. Indeed, segmentation introduces the notion of object into the coder. Thus, the development of functionalities, such as the tracking of objects, becomes a natural extension of coding algorithms. These functionalities come into the framework of the new objectives set by the MPEG-4 standard for video coding.