A Novel Method for Embedded Text Segmentation Based on Stroke and Color

In this paper, a novel method for embedded text segmentation is proposed. The basic idea of our method is based on two properties of embedded texts: a) the color of text pixels is subject to gaussian distribution, b) the locaal part and the global part of embedded text shares the same color distribution. Inspired by this two characteristics, we develop a two-step text segmentation approach: in the coarse segmentation step, a 1-D gaussian function is adopted to model the color distribution of text pixels. To get the model parameters, a stroke operator is utilized to extract confident text region, and then a heuristic process is developed to estimate the parameters. The coarse segmentation can be carried out by the color model. In the noise elimination step, a color distribution homogeneity based method with connected omponent analysis is introduced. Preliminary experimental results show that our method performs well on complex background.