Active Bayesian Mixture Learning for Image Modeling and Segmentation using Lowlevel Features

Gaussian mixture models (GMM) have been shown an effective tool for image representation and segmentation. However, several issues related to GMM training for image modeling have not been adequately resolved such as the specification of the number of mixture components and the increased complexity for images of typical size (e.g. 256 times 256). We present an approach for GMM-based image modeling employing an incremental variational algorithm for Bayesian mixture learning that automatically specifies the number of mixture components. Moreover, we integrate the method in an active learning framework which allows to gradually build the GMM using only a small fraction of the image pixels.