In this paper we describe a robust approach to segment text from color images. The proposed approach mainly includes four steps. Firstly, a preprocessing step is utilized to enhance text blocks in images; Secondly, these image blocks are split into connected components and most of them are eliminated by a component filtering procedure; Thirdly, the left connected components are merged into several text layers, and a set of appropriate constraints are applied to find the real text layer; finally, the text layer is refined through a post-processing step to generate a binary output. Our experimental results show that the proposed approach has a good performance in character recognition rate and processing speed. Moreover, it is robust to text color, font size, as well as different styles of characters in different languages.