Reconstructions faciales à partir d'images tridimensionnelles de crânes humains par recalage et modèle déformable pour l'identification de personnes

This thesis addresses the difficult problem of facial reconstruction from three-dimensional (3D) images of human skulls, considered as unknown. We present an original 3D computerized facial reconstruction method based on recent techniques of medical imaging, where the main objectives are the processing speed and the elimination of the subjectivity by leaning particularly on mathematical criteria for evaluating the results. Our method is based on a model deformable approach constrained by the knowledge of soft tissues thicknesses, at a certain number of characteristic landmarks. It is based on known cases of skull / skin couples. The facial reconstruction is obtained in two main stages. A first registration step of the skull of the known head on the skull of the unkown head, and an estimation of the unkown skin is deducted from this registartion and anthropometric information about the skin thicknesses at different points of the surface. Deformable models adapted to the problem, integrating these constraints, have been used for that. This method has been tested on thirteen sets of skull/skin data, extracted from 3D-MRI of individual heads of children and adults. For each facial reconstruction, two skull/skin sets corresponding to a reference head and a head with unknown skin are selected according to similar anthropological characteristics. By this method, encouraging results are obtained. The reconstructed shapes remain faces, visibly acceptable, and are relatively close to the real faces.