Image Segmentation Using Multi Dimensional Co-occurrence Matrix Based on Multiple Information

Abstract

This paper describes a image region segmentation algorithm : each pixel is segmented, based on the information of the regions located on both above and below the pixel. As an example for the feature value of each target pixel, we utilize the average and the standard deviation of the intensity value of the neighboring pixels, which belong to upper and lower region of the target pixel. Then each target pixel will be projected into and analyzed in a feature space called multi dimensional co-occurrence matrix : the feature space is, conceptually, composed of two pieces of subspace, i.e., one spanned by upper region features and the other by lower ones. In the case of two feature values, the subspace is represented as a two dimensional (2-D) plane, and, the target pixel is projected as a kind of connecting segment between the two subspace planes in the feature space. After having excluded boundary pixels, while fitting 2-D Gaussian distributions, we classify the target pixels into homogeneous regions based on a criterion : segments corresponding to inner pixels have gradients of nearly zero, and those to boundary pixels have much larger gradients.