A method is proposed to segment text from complex background for recognition task. First we estimate the skew of text line and rotate it to the horizontal one. Second, a sampling method based on color edge is proposed to get portion of the text pixels and non-text pixels. Then, k-means algorithm is used to partition the sampled pixels into k clusters. Third, we take the centroid and standard deviation of clusters to segment text image into k binary images. Finally, stroke-filtering is used to select correct one from k binary images. Experiments result shows that the proposed method performs much better than the previous methods.