Abstract:

Signature is an essential authentication means for document approvals and transactions. Online handwritten signature verification has become a key issue in information security with the development of electronic transactions. The principle of signature verification is, supposing the signature of a person is unique and stable, to compare the image, stroke, pressure and velocity information of an inspecting signature with the true signature samples. Thus, the signature can be verified or not. Dynamic time warping (DTW) algorithm is a common algorithm for checking the verification of signatures. It calculates the distance between an inspecting signature and its true signature to judge the verification in the nonlinear time alignment. However, it is hard to determine the threshold, which is related with the individual writing habits. So, it is difficult to do statistics or training. The optimized DTW algorithm, which is based on statistical model of characteristic points, views signature’s characteristic points as a certain probability distribution. Through statistical models, the probability less than the threshold of the feature point is regarded as illegal signature. In this algorithm, the threshold is independent from the individual characteristics. However, this method still needs a lot of a certain person’s true signature samples to train for statistics. In this paper, an improved DTW algorithm based on statistical models is proposed, in which a small amount or even one true signature can be authenticated, and its validity is proved by using this method.

Key words: signature verification, online, DTW, statistical modeling