Online Signature Verification Based On SVM And One-class SVM

With the rapid development of electronic information technology, people has entered the information age, then how to not only identify personal identity quickly and accurately, but also to ensure the security of personal information is an urgent problem. Among numerous solutions to solve this problem, biometric identification technology shows great advantages and prospects. As a member of biometric identification technology, online signature verification is the technology that the personal signature is taken as a biometric feature for identification. Compared with other biometric methods, online signature verification is most likely to be accepted in different situations and more convenient, so it is an ideal personal identification method. Therefore, this article main research the online signature verification technology.The procedure of online signature verification typically includes a series of process such as data acquisition, preprocessing, feature extraction, pattern matching. In this paper, each process has been a systematic and in-depth research, and our main contributions focused on the following parts:First, we propose a new global feature extraction method. Because the global features reflect the signature's overall information and the scale coefficients of discrete wavelet transform reflect the signal's general trend, so based on this, we implement multi-scale wavelet transform on local features and later extract mean and standard deviation on scale coefficients as global features.Second, we deal with the online signature verification problem as a two-class classification problem and propose an online signature verification system based on multi-level matching, which complete the signature verification through three matching stages. In the first matching stage, use the signature time to judge roughly. In the second matching stage, with the advantage of information fusion technology, we fuse global features and local features to match and propose two ways to fuse. In the third matching stage, signature's approximate and detailed information are combined to reflect the similarity between the signatures and use the SVM model to do the final mach.Third, we deal with the online signature verification problem as a one-class classification problem and propose two ways to solve this problem: the first method is the direct use of one-class classifier, so through changing the distance measure form of the kernel function, we design an online signature verification system based on One-class SVM classifier fusion. The second way is to manually generate the negative samples of signature according to the only authentic signatures, then use the binary classification method to classify.In this paper, our research is based on publicly signature database SVC2004. Through a series of experiments, the EER of an on online signature verification system based on multilevel matching is about 3.50%, which is a good verification result and proves the effectiveness of this system. And when the online signature verification based on one-class signature, one way is directly using the one-class classifier and the EER of the system reaches 8.18%, Another way is to artificially produce negative samples and the EER is 6.50%. Experimental results show that in the online signature verification process, the implementation of binary classification has better performance than the implementation of one-class classification.