On-line signature verification based on Gaussian mixture models
Yang Lou, Mandan Liu · 2017
In this paper for on-line signature verification, wavelet packet analysis will be used to extract dynamic local features, combining global features to keep distortionless in signature data. More importantly, in order to overcome shortcomings that the traditional expectation maximization algorithm seriously depends on parameters initialization and easily falls into local optimum when used to train Gaussian Mixture Models, we first employ an improved Splitting-EM algorithm based on Bayesian Ying-Yang learning system to train Gaussian Mixture Models. Splitting-EM algorithm can search for optimal number of Gaussian components so that a unique, user-dependent signature model can be established to ensure a better approximation. Experiments show that the verification accuracy based on wavelet packet analysis to extract features and Splitting-EM algorithm training Gaussian Mixture Models reaches 95.8%, which is a satisfactory verification result.