An Active Search Method for Local Individual Features in Off-Line Signature Verification

摘要

We propose a new method of off-line signature verification that extracts individual "peculiarities" from parts of a signature and uses them for verification. In the registration process, reference signatures are compared with each other for correspondence estimation, and multiple areas exhibiting strong individual peculiarities are extracted. In the verification process, correspondence estimation is first performed, then verification is performed with the extracted local features. We used multiscale regularization for correspondence estimation, and a genetic algorithm for extracting areas that exhibited strong individual peculiarities. As measures of the degree of individual peculiarities we used the magnitude of the deviation as determined from repeated signatures and the degree of complexity of the pen stroke. A verification experiment involving 20 subjects and their Japanese signatures (a total of 800 signatures) produced a best-case average verification error rate of 2.63%.