A Multi-Scale Lip Segmentation Method Based on Markov Random Field

Abstract

Abstract-In order to perform lip segmentation in a lipreading system, we propose a new method based on maximum a posterior and Markov random field (MAP-MRF) framework to statistically model observed images of different texture areas in this paper. First, we establish a multi-scale model to capture the characteristics of each sub-region in the image and use the tree structure in the wavelet domain to calculate the probability of tree nodes at different scales. Thus, the number of layer can be considered as one segment cluster. Then, we utilize MRF to translate the lip segmentation problem into labeling optimization issue. Finally, the Bayesian criteria and the extended expectation maximum (EM) algorithm are applied to estimate child node parameters. The experimental results of this method are more robust than the traditional iterative condition model (ICM).