Expectation-maximization framework for rock textures segmentation

Image clustering can be viewed as a segmentation problem in which small image patches are grouped together based on their features. Rock texture segmentation is a challenging task since the texture is often nonhomogeneous. In this contribution, the new EM (expectation-maximization) rock textures segmentation framework EMRT is proposed. EMRT has two phases, in the first phase the image is divided into small patches then color and texture features are extracted for each patch. We perform EM clustering of the features and then map the clusters back to the image domain to construct segmented image or subimages. The cost function of EM clustering is based on maximum-likelihood (ML) estimation to fit the feature data to a Gaussian mixture model (GMM). In the second phase we use edge detection techniques and morphological operations on each subimage for refinement to define the final segments. A qualitative comparison of EMRT with the traditional approaches such as fuzzy C-mean, and Kmeans on a database of rock images is presented. We proved that the EMRT provides the highest quality segmentation compared with the other approaches