A Comparative Study on EM Algorithms for Color-Texture Image Segmentation

Expectation-Maximization (EM) algorithm has been thoroughly studied in the maximum likelihood estimate of model parameters for statistical learning. Albeit EM algorithms are exploited to the nature of a variety of problems, they are commonly faced with operational difficulty in practice, due to its convergence of local maxima. The actual performance of different EM variants are seldom evaluated to resolve the same application-specific problem, for example image segmentation. In this work, we have conducted a comparative study on different EM variants. To more visually compare them, we employ the EM variants into a color-texture image segmentation algorithm. We first evaluated the effectiveness of several EM variants using the log-likelihood and Bayesian Information Criterion on the image data. Then the EM variants are used for color quantization in the framework of the color-texture segmentation algorithm to assess the performance of them. The local maxima problem of the EM algorithm is also studied by the image segmentation results.