Text localization in complex images using structural information and its application to recognition of slab information numbers

Abstract

Content-based image indexing is to label image based on its content. Such as color, texture, shape, face, text, and etc. can be kind of image contents. Especially, because the text has been considered to be easily localized and recognized compared to other image contents many researchers have studied about the general text information extraction algorithm in images actively. However, the more the ambiguity between the text and background, the text information extraction algorithm also has to be designed more complexly likewise other contents. The unfortunate fact here is that the complexly designed algorithm does not always mean increase of performance. Taking these fact into account, researchers have put their energy in eveloping generic text information extraction system. In spite of these efforts, because (1) it is difficult to find features that enhances only the text itself and (2) there are various forms of texts (characters), and (3) actually it is impossible to include all texts forms into the training and test set, it is still difficult to design the generic text information extraction system. The target data in this thesis show mostly low quality affected by such as lighting condition, unwanted shadow, low quality text itself, and so on. To overcome these difficulties, the presented text localization method (1) finds top and bottom boundaries of text candidates using geometrical property, (2) selects a true text, and (3) find left and right boundaries of text rectangle through the text model and the designed fitness function. Developed text localization method simultaneously segments characters without any additional complex segmentation stage. Comparison with other existing algorithms conducted andexperimental results show that the developed system is superior to others and can be successfully applied in the fields.