Generalized Gaussian mixture Conditional Random Field model for image labeling

Abstract

This paper proposes new potential functions for Conditional Random Fields (CRF) in image labeling framework based on generalized Gaussian mixture Modeling (GGMM) of the potential functions. Laplacian mixture potential functions have previously been applied to CRF. However, Laplacian potentials fail to capture data characteristics where data fluctuations happen very smoothly; so that they even give rise to induction of atypical results due to erroneous modeling of data. Having an additional shape manipulation parameter, generalized Gaussian mixtures (GGM) can model data characteristics and fluctuations precisely. In this paper, we propose to deploy GGM in the CRF framework to formulate the potential functions. Expectation maximization (EM) technique is used to estimate GGM parameters. Belief propagation and stochastic gradient descent algorithms are utilized for CRF inference and training, respectively. We show that proposed GGM feature functions effectively improve labeling accuracy of nature images in comparison with Laplacian mixtures. Qualitative labeling results show that the proposed framework performs well particularly for labeling simple even backgrounds where the Laplacian counterparts impose irregular outcomes. That is, despite Laplacian mixtures, GGM-based feature functions can correctly model smooth image color and texture variations.