Font Size: a A A

A Study Of Off-line Chinese Signature Verification

Posted on:2007-05-01Degree:DoctorType:Dissertation
Country:ChinaCandidate:Z H WuFull Text:PDF
GTID:1118360242461950Subject:Computer system architecture
Abstract/Summary:PDF Full Text Request
As an identity authentication technology, signature verification is used in many kinds offield such as many types of document, business contract and bank cheque et al. Up to the present,researches about automatic off-line signature verification mainly are verifying random forgeriesand simple forgeries. Especially for simulate forgeries, there are many theories and applicationproblems have to be resolved. So, there are great theoretical significance and practical value tocarry out the research of automatic of?ine Chinese signature verification.Around three basic research contents of pattern recognition: data extraction and preprocess,feature extraction and selection and matcher, several key technologies involved by automaticoff-line signature verification are discussed summarily. Then, in term of the basic theories ofnoise elimination and the key problem that should been concerned when selecting the method ofnoise elimination, and according to the characteristic of pixels layout in line, a horizontal lineclearing method is presented. Compared to generic noise filters, this method can greatly decreasein?uence on the origin signature image and is easier to achieve.Calculation complexity and correct rate of classifying are the two main aspects concernedby us when selecting feature extraction method. In connection with the classifying of Chinesesignature, to address better re?ecting the characteristics of signature, how to use global featuresto compose the feature vector is discussed. Based on these, two kinds of clustering methods, oneis based on Mahalanobis distance and C-means method, the other is kernel clustering based onfeature-weighting, are presented.Signature segmentation is a difficult task in preprocess stage. The quality of segmentationwill greatly affect the verification result. A simple, low computational cost and robust segmen-tation method is proposed by means of having successful experiences of strokes extraction forhandwritten Chinese character and taking into account the characteristics of signature verifica-tion. The algorithm presented has a low time and space complexity. Then, a segments sorting andsimilarity degree computation method are presented on the basis of Euclidean distance. Based onthese, a classifier composed with template matching and RBFNN (Radial-Based Function NeuralNetwork) is described.The quality of vector quantization is a crucial factor when using DHMM (Discrete Hid-den Markov Model) for pattern recognition. In view of this, the feature vectors extracted fromsegments are grouped into two groups according to their physical meaning and each group isquantized respectively. In the vector quantization method, using multi-codeword to represent avector is used. In connection with these two optimization methods of vector quantization, thelearning algorithm of HMM (Hidden Markov Model) is adjusted slightly. Compared to grid method, the method based on stroke segmentation presented in this thesis reserves the layout ofsignature strokes and can extract a better observation sequence. A learning method of DHMM'sparameters which has both excellence of GA (Genetic Algorithm) and Baum-Welch is proposed.The optimized DHMM improves the correct rate in the sense of statistics.To address the verification of simulated signature, the extraction method of more detailedfeatures is discussed. An efficient tracing algorithm for the big close-ring of signature is pro-posed. Compared to the searching method of tracing route, the tracing algorithm proposed de-creases the time complexity and space complexity greatly. According to the layout of pixels andthe pixel's direction come from, this tracing algorithm chooses the next pixel, which ensures thetrace process only needs one time and the tracing order consist with the certain order. On thebasis of edge tracing, the extracting method of secondary features is discussed by using waveletpacket transform. The HMM classifier and SVM (support vector machine) classifier are com-pared by using the features extracted. The comparison result shows HMM is better than SVM inour experiments.At last a cascading combination system composed with three classifiers is built by combin-ing an appropriate application. The verification system solves the problem of genuine signaturecollection. Each classifier verifies one type of forgery. These classifiers are sorted from rough-ness to fine, which decreases the verification time effectively. The promising results of test usinggreat scale actual data indicate the verification system presented in this thesis can finishes theautomatic off-line signature verification task very good.Through the study on automatic of?ine Chinese signature verification, we achieve sometheory and applied results, and establish the theoretical and methodological basis for the appliedsystem's research.
Keywords/Search Tags:Feature Extraction, Template Matching, Hidden Markov Model, Classifier Combi-nation, Signature Verification, Signature Segmentation
PDF Full Text Request
Related items