Research Of BP Network Signature Verification Using Statistics Features And Wavelet Features

With more and more attention paid to the issue of identity verification in network grading system, the signature verification technology based on biometric features, which is non-invasive, easily remembered and applied to a wide range, has broad application prospects in the network grading system.The usefulness information is removed from the original signature information by pretreatment and the signature data is converted to a form suitable for extracting the signature features. The pretreatment of this paper is composed of removing zeros removing noises, normalizing in direction and so on. The verification of true signature and the forgery signature is realized by extracting the signature features and using the BP neural network classifier.Extracting signature features is completed by extracting the statistic features from speed, time and shape of the signature, and extracting the wavelet features composed of the high frequency coefficients by decomposing the horizontal and vertical shift of signature with DB6 wavelet. Through comparing the true signatures and the forgery signatures repeatedly, the conclusion of comparison shows that the statistics feature vector composed of 15 statistics features and the wavelet feature vector composed of 64 wavelet features are able to represent the identity of a person by being stable for the same person and distinctive for different persons.The improved BP network classifier is used; both the true signatures and the forgery signatures are put into the network to get trained, and the connected weights between the neural units that have been trained successfully are saved. Designing the network structrues, selecting the mode of training and improving the standard BP algorithm are the works of this paper. The number of neural units of the hidden layer is determined by the experimental formula combined with the statistics features and the wavelet features. The batch mode of sample training is selected after comparing with the single-sample mode. The problems of a local minimum and a slower convergence speed are solved by introducing the momentum factor and variable step length, as the standard BP network is getting trained.Three methods of verification, which are the BP network single-level verification using statistics features, the BP network single-level verification using wavelet features, and the BP network two-level verification using statistics features and wavelet features, are compared by making experiments of the true signature samples and the forgery signature samples. For the true signature samples, the false rejection rates named FRR of the three methods respectively are 8.0%, 6.0% and 6.0%. For the first-class random forgery signature samples, the false acceptance rates named FAR of all the thress methods are 0%. For the second-class random forgery signature samples, the FAR of the thress methods respectively are 4.4%, 2.4% and 0.4%. For the skilled forgery signature samples, the FAR of the thress methods respectively are 19.2%, 15.6% and 2%.The result of experiments shows that the two-level verification mode with a lower FAR is obviously superior to the single-level mode.