Texture Classification in Microstructure Images of Advanced Materials

In this paper we discuss texture classification of microstructural images used in the development of advanced materials. Different materials or different phases of a material in a sample lead to different textures in images of that sample. We investigate the application of the expectation-maximization/maximization of the posterior marginal (EM/MPM) algorithm proposed in [1] for automated texture-based segmentation of these microstructure images. We propose a new method to initialize the parameters used for the observed image under a Gaussian model. In addition, we examine the use of 8-nearest-neighborhood system for segmentation to locate objects edges more precisely. The experimental results demonstrate the adaptability of the EM/MPM algorithm with appropriate choices of model parameters and neighborhood system. Furthermore, we demonstrate that the use of 8-nearest-neighborhood system can provide smoother segmentation in edges.