Online Handwriting Signature Verification Based on Template Clustering

Template-based and template-matching methods are commonly used in online signature verification, and it is very important to select a suitable template signature. Generally, template signatures are chosen from the enrollment signatures, but when the number of enrollment signatures is large, the computation time is long, and some enrollment signatures are not representative. This paper submits a two-level signature verification method based on template clustering. Firstly, a similarity score matrix is obtained by comparing the enrolled signatures one by one, and the median of all similarity scores is taken as a clustering threshold. Then the signature threshold clustering algorithm is used to cluster the enrolled signatures. In each signature cluster, the signature with the highest aggregate score is chosen as the signature representative of the signature cluster, and the other signatures in each signature cluster are used as the second-level template signatures. In each one-to-one signature matching process, a fusion feature weighted by the similarity distance and time-speed ratio of the fusion curve is considered as the basis for signature similarity evaluation. Finally, on two open signature datasets, SVC2004 Task 1 and SUSIG Visual, different numbers of genuine signatures are selected as training set for template clustering. Remaining genuine signatures and skilled forged signatures are used as test set for experimental comparison. Depending on the number of training signature samples, our best equal error rate (EER) on two datasets is 2.50% and 0.83%, respectively. The experimental results test the effectiveness of the proposed method.