Information fusion systems are mainly composed from mathematical tools allowing to realize data representation and combination. The aim of these systems can be expressed as a decision problem on the truth or plausibility of a proposition based on several information coming from different sources. Fusion try to manage the characteristics of the sources taking into account the information imperfection (inaccurate, incomplete, ambiguous, uncertain, etc.) and the redundant aspect, the complement and the conflictual aspect of information. Fusion systems concerned by this thesis have the ability to integrate the expert knowledge in their treatments. They are called cooperative fusion systems. Since these systems are trying to associate experts, it is important to provide to the users some informations that help them to better understand the fusion process. Such systems have many parameters that must be adjusted. These parameters have an important impact on the quality of the obtained results. One of the major problems associated to information fusion systems concerns the evaluation of their performance. A pertinent evaluation will allow to improve the quality of the fusion, to improve expert/system interaction and to better adjust the parameters of the system. Generally, the evaluation of such systems is made in the ouput of the processing chain by a global evaluation of the results. But it does not allow to know the precise subpart of the treatement chain that requires an adjustment of its parameters. Another difficulty releases in the fact that a complete ground truth of the result is not always available, which complicates the performance evaluation task. The application context of this work is the interpretation of 3D images (tomographic images, seismic images, synthetic images, ...). In this context, a local evaluation of the information fusion systems has been implemented. The approach has shown its interest in the efficient adjustment of parameters and the cooperation with expert.
