Unsupervised Segmentation of Colored Texture Images by Using Multiple GMRF Models and a Hypothesis of Merging Primitives

Abstract

Images of outdoor scenes involve many textured areas. In terms of coding and recognition, it is important to extract textured areas using not only color but also various texture features. Texture features are represented at segment level rather than by individual pixels, hence it is difficult to use such features near area boundaries. In this study, Gaussian Markov random field modeling is employed with multiple independent parameters. In so doing, preprocessing is performed to minimize correlation among color planes, and GMRF parameters are estimated using hypothetical merging, so that texture features can be used near segment boundaries as well. As merged segments grow in size, more precise GMRF models are used which results in images with smooth boundaries. © 2000 Scripta Technica.