Unsupervised image segmentation utilizing penalized inverse expectation maximization algorithm

Abstract

This work is on accurate segmentation of images using local image characteristics. An appropriate Gabor filter with customized size, orientation, frequency and phase for each pixel is selected to measure the image features. A new image property called phase divergence is introduced to select the filter size. Brightness, color, texture and position features are extracted for each pixel and the joint distribution of these pixel features is modeled by a mixture of Gaussians. A new version of the expectation maximization (EM) algorithm called Penalized Inverse EM (PIEM) is formulated for estimating the parameters of the mixture of Gaussians model. Furthermore, we determine the number of models that best suits the image based on Schwarz criterion. The performance on the Berkeley segmentation benchmark proves the efficacy and accuracy of the proposed method.