Gauss Mixture Image Classification for the Linear Image Transforms

Abstract

Gauss mixture models are commonly used in image classification due to their analytical tractability and robustness. When the feature vectors are formed as the coefficients of a linear image transform, the underlying mixture components are not necessarily Gaussian, in which case there is no guarantee that the Gauss mixture model (GMM)-based clustering algorithms can capture the mixture components. In this work, we train an unbalanced tree-structured GMM-based classifier to reduce this problem. We derive and apply a parameter-independent test to determine the number of mixture components in any given tree node. The classifier tree is grown only in the regions with multiple mixture components.