Research And Application Of Printed Font Recognition Based On Texture Analysis

Printed font recognition plays an important role in OCR system. Font recognition not only can make the OCR system more roboticized, but also can provide font information for character identification so as to reduce character template matching time and improve Recognition Rate. So far, the researches on font recognition are time-wasting in feature extraction, or strict in printed quality of those document. So, these methods can not be applied in OCR system.In this paper, research is carried out with 4 frequently used Chinese fonts in newspaper, viz. Song Style, Kai Style, Fangsong Style, Hei Style. Different fonts are regarded as different textures, then font is identified through texture analysis with multi-channel Gabor filters. Main works based on the newspaper samples is just as the following: (1) Texture block is designed to embody more font characters. (2) Filter orientation is optimized with Genetic Algorithm to make the angle set subtler. Then the multi-channel Gabor filter may extract better features. (3) Because the distribution of the font features is multi-apex, subsection linear-classifier got by dynamic clustering arithmetic is employed to better classify the fonts. As far as the samples of Guangming Daily and People's Daily are concerned, features are extracted by Gabor filter with coarse angles, and fonts are classified by simple linear-classifier, the Average Font Recognition Rate(AFRR) is 76.31%. But the AFRR reaches 95.62% after the above works.Finally, this research is applied to practical system. And some adaptions are made in recognition velocity and methods. Those adaptions make the AFRR increase to 97%. And method of individual character font recognition based on guidance font is used to identify the non-dominant font in a document. The results of experiments show that the AFRR is 98.6%.