Unsupervised segmentation of natural images based on statistical modeling

Abstract

A novel unsupervised scheme for natural image segmentation is proposed aiming to acquire perceptually consistent results. Firstly, comprehensive visual features besides raw color values are extracted, including spatial frequency, contrast sensitivity, color deviation, and so on. Secondly, high correlations among visual features are reduced via principal component analysis (PCA) and the raw image pixels are then converted to a collection of feature vectors in a multi-dimensional feature space. Thirdly, the Gaussian mixture model (GMM) is employed to approximate the class distribution of image pixels and an improved expectation maximization (EM) algorithm is introduced to estimate model parameters. Finally, segmentation results are obtained by grouping of pixels based on the mixture components. Experiments are conducted and the results demonstrate that, compared with existing techniques, the proposed scheme can acquire more perceptually consistent results.