Improved Heterogeneous Gaussian and Uniform Mixed Models (G-U-MM) and Their Use in Image Segmentation

Abstract

Recently, the combined Gauss Mixture and Uniform Distributions Mixture Model, shortly Gauss-Uniform Mixture Model (G-U-MM) was proposed to better relate to the nature of a complex distribution and to simplify the characterization of processes that need too many Gauss functions in a standard Gauss Mixed Model (GMM). For a reasonably large class of images, the Gauss-Uniform distribution mixed models are easier to apply than the GMM models because the former ones produce signicantly smaller numbers of elements in the mixture. The method has solid mathematical foundation and might be better related to the processes of image segmentation performed by humans. In addition, while computationally simple, it produces remarkable results. We discuss supplementary reasons for the use of the G-U-MM heterogeneous models in image segmentation and improve the previously presented algorithm of segmentation by removing the possible confusion between sections of Gaussian distributions and intervals of uniform distribution. Consequently, the approximation precision of the histogram and the segmentation are improved. Several examples illustrate the algorithm performance.