A clustering strategy for touching characters in Korean and English printed text segmentation

This paper proposes segmentation method in mixed Korean and English printed text which contains touching characters using clustering strategy. At the first step, a vertical projection of image text is determined, and clustering process performed on it. Then the cluster with the smallest mean value used as candidate segmentation point. This process will produce candidate bounding boxes. Furthermore, they should be verified whether according to Korean or English characteristics otherwise they will be splitted or merged each others. The merged process could be done based on Korean vowel characteristics since Korean alphabet consist several symbols, while splitted process could be done by local vertical projection clustering. The proposed method gives 99.36% correct segmentation rate in un-touching characters and 99.25% in touching characters. This result shows that the proposed method using clustering strategy is very effective for touching problem in mixed Korean and English printed text. Besides, it also improves the speed of segmentation process, because the method does not need a character recognizer to verify bounding boxes.