Study On Uyghur Handwritten Signature Recognition Based On Multiple Feature

In order to solve the problem that the personal information is often exposed in the modern society,researchers from different regions are paying attention to the individual biometric technology which cannot be easily stolen.As a kind of stable biological behavior characteristics of human body,handwritten signature has the characteristics of non-invasion and easy to access in the aspect of personal identity authentication.Therefore,handwritten signature recognition technology has been widely used in many fields and has played an important role,such as bank service window deposit and withdrawal,signing express mail,communication card,signing a contract and so on.In order to meet people's needs for the protection of personal information in modern society,it has great practical value and important significance to find a more secure,universal,convenient,fast,reliable and practical means of identity authentication.This paper firstly introduces the background of the research of signature recognition?the practical application value?significance and overview,and then briefly introduces the status quo and the results of the research of signature recognition at home and abroad.Then combined with the status quo of Uyghur handwriting signature research,step by step to introduce our proposed method of pre processing?feature selection and extraction?classification matching method.Finally,two kinds of training samples were used to experiment and the experimental results were analyzed.In data acquisition and preprocessing phase,the original signature sample base of Uygur handwritten signature were collected,and then it is carried on the signature image graying,smooth denoising,binarization,normalization and contour extraction.In the feature selection and extraction phase,128-dimensional local center point features,112-dimensional ETDT features,30-dimensional texture features,10-dimensional Zernike moments and 12-dimensional shape features were extracted from each normalized binary signature image and contour signature image separately.Among them,the local central point features and ETDT features were combined to form a new high-dimensional statistical fusion feature,and then texture features,Zernike moments features and shape features were combined to form a new fusion features.In the feature classification decision part,for the local center point feature?ETDT feature and the fusion of these two features,K-NN was used to sort firstly,and then it has been used the absolute distance,Euclidean distance,chi square distance and Cosine distance similarity method to find the closest and most similar category.For the texture features?Zernike moment features?shape features and the fusion of these three features,BP neural network was used for training and classification.The experimental data consisted of 100 volunteers(each volunteer have 20 signature samples),with a difference in educational level and age,a total of 2,000 signature samples.In this paper,two training modes(1600 samples and 1000 samples were used as training samples)were used for the experiment and the results were analyzed.In the experiment,when the number of training samples Was selected 1600,128 dimensional local center features and 112 dimensional ETDT feature respectively were used absolute distance and Cosine similarity measure for classification,and it was obtained 96.35% and 97.45% of average recognition rate.In order to solve the problem that the recognition rate using only a single feature is not high enough,the local central point features and ETDT features were combined to form a new high-dimensional statistical fusion feature,and the similarity distance measure method was used for the classification experiments.The recognition rate of fusion feature was higher than 0.9% and 0.5%,respectively,and the average recognition rate was about 98.35% and 96.85%,separately.In addition,for the 30-dimensional texture features,10-dimensional Zernike moment features and 12-dimensional shape features,BP neural network was used for training and classification,and it was get 95.45%,97.80% and 93.39% of average recognition rate respectively.In order to complement these single features mutual defects,three kinds of features,the texture features,Zernike moment feature and shape features were combined to form a new fusion feature,and BP neural network was used for classification and recognition.The recognition rate of the system is higher than the recognition rate of the single feature about 2.55%,0.2% and 4.61%,respectively,and the highest recognition rate was up to 98%.