Skin detection or segmentation is considered as an important preliminary process in a number of existing systems ranging over face detection, filtering Internet images, and diverse human interaction areas. Nevertheless, there are two skin segmentation challenges: the pattern variability and the scene complexity. This thesis is devoted to define a new approach for modeling the skin probability distribution. ln the aim of dealing with the skin detection problem, we investigate the models of probability trees to approximate skin and non-skin probabilities. These models can represent a joint distribution in an intuitive and efficient way. Hence, we have proposed three main approaches to seek a perfect tree model estimating the skin probability distribution: (1) the model of dependency tree that approximates the skin and the non skin probability distribution together, (2) the mixture of trees' model, and (3) the combination of trees' model. The first proposed model is based on the optimal spanning tree principle combined to an appropriate relevant criterion that we have defined. The contribution takes into account both the interclass and the intra class between skin and non skin classes, and the interactions between a given pixel and its neighbors. The rationale behind proposing the second model is that in sorne cases the approximation of true class probability given by an optimal spanning tree (OST) is not unique and might be chosen randomly, while this model will take the advantages of the useful information represented on each OST. The mixture of trees' model consists in mixing the structures of the OSTs and their probabilities with the aim of seeking a perfect spanning tree. This latter emphasizes the dependencies' degrees of data, and approximates effectively the true probability distribution. Finally, the third model is defined to deal with a particular kind of multiple OSTs. This model is a parallel combination of different classifiers based on the OSTs. A mathematical theory, proving and specifying the appropriate approach to be used (mixture of trees or combination of trees) depending on the considered OSTs' kind, is presented in this thesis. In addition to experimental results, on the Compaq database, showing the effectiveness and the high reliability of our three approaches.
