Mixture-State Document Segmentation Using Wavelet-Domain Hidden Markov Tree Models

Abstract

In this paper we introduce a mixture-state document segmentation method based on wavelet and the hidden Markov tree (HMT) models. First we propose a three-state HMT segmentation method that is similar to those in the reference [1]. Then through comparing the difference weights to the three-density Gaussian mixture distribution of different textures, we find that background, text and image can be well approximated respectively by one-state and two-state and three-state HMT models. Then we get a new segmentation method, mixture-state HMT segmentation. Experiments with scanned document images indicate that the new approach improves the segmentation accuracy over the raw segmentation in [1].