Spatially-Coherent Segmentation Using Hierarchical Gaussian Mixture Reduction Based on Cauchy-Schwarz Divergence

摘要

Gaussian mixture models (GMM) are widely used for image segmentation. The bigger the number in the mixture, the higher will be the data likelihood. Unfortunately, too many GMM components leads to model overfitting and poor segmentation. Thus, there has been a growing interest in GMM reduction algorithms that rely on component fusion while preserving the structure of data. In this work, we present an algorithm based on a closed-form Cauchy-Schwarz divergence for GMM reduction. Contrarily to previous GMM reduction techniques which a single GMM, our approach can lead to multiple small GMMs describing more accurately the structure of the data. Experiments on image foreground segmentation demonstrate the effectiveness of our proposed model compared to state-of-art methods.