Improvement of On-line Signature Verification Based on Gradient Features

This paper proposes a new on-line signature verification technique which employs gradient features and a pooled within-covariance matrix of training samples not only of the user but also of the others. Gradient features are extracted from a signature image reflecting the velocity of pen movement as the grayscale so that both on-line and off-line features are exploited. All training samples of different signatures collected in design stage are pooled together with the user's samples and used for learning within-individual variation to reduce required sample size of the user to minimum number. The result of evaluation test shows that the proposed technique improves the verification accuracy by 4.9% when user's sample of size three is pooled with samples with others. This result shows that the samples of different signatures are useful for training within-individual variation of a specific user.