Character image extraction from a scene image using color information and multiscale analysis

In this paper, a method for extracting a character image from a scene image using color information and multiscale analysis is proposed. In scene images, it is difficult to separate a character image from the background accurately because of noises and the change in the lighting condition. In the proposed method, first the number of colors is reduced by a color clustering technique and several binary images are generated from the clustering results. Meanwhile, a character region is estimated by using the edges detected from the image. In the estimated region, a global structure of the character is obtained. Finally, the binary image that represents the character is obtained by unifying the information of the global structure and the binary images. The effect of the proposed method is evaluated with the experimental results using the ICDAR2003 datasets.