Study Of Off-line Handwritten Chinese Characters Recognition Based On Binary Tree SVM With Stroke Density

Nowadays off-line handwritten Chinese characters recognition, whichhas comprehensive application prospects, is one of the research hotspots inpattern recognition field. Off-line handwritten Chinese characters not onlyshare certain common attributes, such as considerable character sets, varioushandwritings, kinds of changeable glyphs and numerous similar Chinesecharacters, but also have a few features such as numerous writing styles,non-standard writing and higher randomness and so on. On the basis ofanalysis about off-line handwritten Chinese characters’ stroke density feature,this thesis does study of combining two methods to recognize Chinesecharacters. The one is binary tree SVM under the stroke density feature ofChinese characters which is used on off-line handwritten Chinese charactersmultilevel recognition, the other one is one-to-many SVM which is used ondetailed classification. Work in following aspects has been done in this thesis:1. Under the study on structure feature and statistic feature of Chinesecharacters, definitions about the whole stroke and detailed stroke (horizontal,vertical and oblique) density feature of off-line handwritten Chinesecharacters have been given. They can be used as the basis of rough multilevelclassification on off-line handwritten Chinese characters.2. The category on rough classification of Chinese characters has been set upunder the statistics and analysis on the distribution of off-line handwritten Chinesecharacters’ pixels density feature. According to the division of cough classification’scategories, binary tree structures specific to different classification strategies havebeen constituted, and the training on binary tree SVM has been done. The definitionof similarity which is applicable to cough classification of pruned binary tree SVMand cough classification algorithm have been provided as well. Simulationexperiments show that off-line handwritten Chinese characters multilevel recognitionachieves desired effect.3. On the base of rough classification, the peripheral contour feature of off-linehandwritten Chinese characters and wavelet multi-grid features have been extracted asthe input of SVM detailed classification and recognition. Then SVM one-to-manyalgorithm which is applicable to off-line handwritten Chinese characters recognition has been studied. Simulation experiments show a good recognition result.This thesis chooses the samples of SCUT-IRAC HCCLIB handwritten Chinesecharacters as experimental data and MATLAB R2011a as simulation platform, andthen simulation experiments have been done to test the method of binary tree SVMclassification and recognition under pixels density. Experimental results show thevalidity of this method.