An across-scale fusion approach to segment document image

Segmenting an image into text and picture area is very important for efficiently compressing document images. The paper introduces an across-scale fusion approach to segment document images, which makes use of the multiscale down-sampling bi-level images and a Markov tree model in order to directly calculate the classification based original image to be segmented. The main process of the method is divided into a classification stage and a segmentation stage. In the first stage, we separate picture from text blocks based on multiscale bi-level images. Then we segment the image into the foreground level and the background level in the second stage. The paper describes in detail a multiscale representation of the image, the Markov tree, likelihood computation for classifying image blocks, and a modified bi-color clustering algorithm for segmenting the non-picture area.