This thesis describes the design and realization of a complete processing chain for content based image retrieval (CBIR). The study allows to define some "limited semantics" with respect to the user's satisfaction from the system response. The image is decomposed on visual entities to obtain interactions between them, allowing to reach higher levels of abstraction. We have addressed three points in the chain : reliable region-detection, region characterization and then similarity measure. We have modified a Fuzzy C-means by incorporating the spatial and multiresolution information into the objective function. Therefore, the classification of a given point is forced to follow both neighbors and ancestors in a pyramidal representation. Two methods are proposed which exploit Peano scans to coding region features. The first one is based on a grammatical representation of the pixels neighbourhood called motif. The second method modifies the spectrum before to apply Gabor filters. The image signature consists of a list of visual entities containing features. The similarity measure between two images turns into a graph matching problem. We have elaborated a technique that allows a bidirectional matching from query to target and vice versa. A high priority is assigned to those elements which prefer mutually. Each part of the system is evaluated and tested independently then incorporated into the CBIR application. The evaluation of CBIR in terms of "recall-precision" shows that the proposed methods perform better than classical ones, such as grey level co-occurrence matrix and Gabor filters. To open on further extensions and suggest the generality of our method, the conclusion deals with extending it to the situation assessment in car driving, with limited tuning of parameters.
