Signature alignment based on GMM for on-line signature verification. (May 2017)

• Record Type: Journal Article
• Authors: Xia, Xinghua; Chen, Zhili; Luan, Fangjun; Song, Xiaoyu
• Is Part Of: Pattern recognition. Volume 65(2017:May)
• Page Start: 188
• Page End: 196
• Publication Date: 2017-05
• Subjects: On-line signature verification -- Signature alignment -- Gaussian Mixture Model -- Signature curve constraint -- Dynamic weight factor; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2016.12.019
• Languages: English
• ISSNs: 0031-3203

Abstract

Abstract: On-line handwritten signatures are collected as real-time dynamical signals, which are written on collective devices by users. Since writing environments are always changed, fluctuations can be caused by signature size, location and rotation angle which being various at each inputting. Signatures should be effectively aligned before verification, which can diminish deviations caused by these fluctuations. In this study, we propose a method of signature alignment based on Gaussian Mixture Model to obtain the best matching. In verification, a modified dynamic time warping with signature curve constraint is presented to improve the efficiency. Weight factors are dynamically assigned to features, which depend on coefficient of variation, to improve the robustness. Several experiments are implemente.d on the open access on-line signature databases MCYT and SVC2004 Task2. The best performances can be provided with equal error rates 2.15% and 2.63%, respectively. Experimental results indicate the effectiveness and robustness of our proposed method. Highlights: Signatures are aligned well to reference template based on Gaussian Mixture Model before verification. Efficiency of verification is improved by modified dynamic time warping with signature curve constraint. Distance probability is presented to describe the dissimilarity between test signatures and references. A strategy of dynamic weight factors being assigned to features depend on coefficient of variation.