Research On Off-line Uyghur Handwritten Signature Identification

The handwritten signature is one of the biological features that it has a longhistory, nowadays; signature still plays a very important role in social life. Signatureis widely accepted so that it has widespread application in business, finance, judicial,insurance, communication, office automation and credit cards fields. Therefore, it hassignificant theoretical meaning and great practical value to further study in usingcomputer to recognize off-line handwritten signature.The main work of this paper is divided into three aspects such as, preprocessing,feature extraction and selection, classification judgment. In preprocessing stage, bythe preprocessing of signature images such as grayscale, smoothing, binarization,thinning and normalization carries give an effective representation which paves theroad for a perfect feature extraction.In the feature extraction part, it is extracted density feature,16-dementional and64-dementional local central point feature separately from each signature in thispaper.According to the nature of data and extracted features, it is respectively classifiedwith K-NN classifier, Ordinary distance, characteristic distance, vector distance basedmeasure methods after extracting effective features.The effects on Ordinary distance, Characteristic distance, Vector distance basedmeasure methods to the recognition rates are comparatively analyzed, and the bestmeasure method for the different features are confirmed. The datasets used in thisexperiment which collected different75person’s1500samples (20samples/person) areselected from Uyghur handwritten signature database. According to usage, thesessignatures are divided two parts: training dataset and testing dataset. It is selected4different kinds of combination for training dataset and testing datasets that trainingdatasets are set1200samples,1125samples,750samples and600samplesrespectively, and others are used for testing from these1500samples.When training1200samples, it is obtained96.67%and96%of highestidentification rates with using16-demenntional local central point features plusOrdinary distance and Density features plus Vector distance. In order to furtherimprove the recognition rate, the dimensionality of local central point features areincreased to64, and it is achieved100%of best accuracy on the mentioned all of4different types of dataset combination modes when they are classifying with Featuredistance and Vector distance in this paper.