Image and video segmentation with mixture-based semi-supervised clustering

non disponibile

Image segmentation is one of the fundamental problems in Computer Vision, one that has received countless studies and generated several algorithms and techniques. Albeit simple to state as a problem, partitioning a digital image into multiple regions is an open problem as it cannot be objectively formalised, and there really is no general solution. Thus, general-purpose techniques must be combined with prior knowledge in order to be effective. Analysis of video sequences presents even more challenges due to the intertwined spatial and temporal dimensions, but allows for several interesting inferences about shapes and motions of consistent regions/objects. Recently, a lot of attention has been directed towards injecting prior knowledge into the basic frameworks of probabilistic models in order to “bend” their strong modelling power towards domain-specific solutions. It is in this context that we see the viability of semi-supervised clustering, i.e. clustering under the influence of additional information. In this thesis, we describe an original framework to perform semi-supervised clustering with probabilistic mixture models. These models are tailored to deal with the specific nature of images and video sequences in order to be effective and efficient. To estimate the parameters of the proposed models, we derive a (generalized) EM algorithm with a closed-form E-step and introduce a novel updates scheme that exploits the strengths of our particular formulation. We show several experimental results with known image databases and benchmark video sequences, with quantitative comparisons to other state-of-the-art techniques where possible and relevant.