Isolated Character Recognition by Searching Feature Points and Voting

To utilize a character information in scene images, character recognition and segmentation are required. Character segmentation is essential preliminary of character recognition. However, most character segmentation methods cannot extract isolated characters which do not constitute a character string, characters in a complex equation, touching characters where two characters are connected and so on. Therefore, they cannot be recognized. In this paper, to utilize information of such characters, we propose a novel recognition method based on extracting feature points and voting. The voting algorithm of the proposed method is similar to the generalized Hough transform. The proposed method generalizes the conventional character recognition for character images segmented correctly to one for unsegmented character images. In experiments, the proposed method extracted isolated characters in the face of a clock and touching characters well even if the fonts of a reference image and a target one were different.