Abstract

Model-based approaches have become important tools to model data and infer knowledge. Such approaches are often used for clustering and object recognition which are crucial steps in many applications, including but not limited to, recommendation systems, search engines, cyber security, surveillance and object tracking. Many of these applications have the urgent need to reduce the semantic gap of data representation between the system level and the human being understandable level. Indeed, the low level features extracted to represent a given object can be confusing to machines which cannot differentiate between very similar objects trivially distinguishable by human beings (e.g. apple vs tomato). Such a semantic gap between the system and the user perception for data, makes the modeling process hard to be designed basing on the features space only. Moreover those models should be flexible and updatable when new data are introduced to the system. Thus, apart from estimating the model parameters, the system should be somehow informed how new data should be perceived according to some criteria in order to establish model updates. In this thesis we propose a methodology for data representation using a hierarchical mixture model basing on the inverted Dirichlet and the generalized inverted Dirichlet distributions. The proposed approach allows to model a given object class by a set of components deduced by the system and grouped according to labeled training data representing the human level semantic. We propose an update strategy to the system components that takes into account adjustable metrics representing users perception. We also consider the "page zero" problem in image retrieval systems when a given user does not possess adequate tools and semantics to express what he/she is looking for, while he/she can visually identify it. We propose a statistical framework that enables users to start a search process and interact with the system in order to find their target "mental image". Finally we propose to improve our models by using a variational Bayesian inference to learn generalized inverted Dirichlet mixtures with features selection. The merit of our approaches is evaluated using extensive simulations and real life applications.