Separating characters from graphics is an important step towards automatic document understanding. In this paper, we propose a robust algorithm to separate Chinese characters from graphics. Our approach is based on clustering the feature points in an image. Two remedy procedures are also proposed to solve the problems caused by the thinning process. This will obtain a better localization of feature points and improve the performance of the separation process. Using our algorithm, all Chinese characters can be separated from graphics without regard to the font style or orientation of the character. Furthermore, our algorithm can also handle the serious case where characters touch/cross lines. The proposed algorithm has been successfully tested on several kinds of line drawings, such as land register maps and form documents.