On-line signature verification is a process to determine the genuiness of a signature written by an electronic pen. It has the advantage that dynamic writing information can be used as well as shape information.
This thesis presents the on-line signature verification system based on segment-to-segment comparison. A model-based segmentation method is proposed to obtain stable and globally consistent sementation which cannot be obtained by simple external segmentation methods. An input signature is segmented by the correspondence with the model signature. Therefore, we can get more reliable matching between segments in the sense of shape similarity. On the other hand, a selective utilization method of the features is proposed. This approach is motivated from the observation that a skilled forger can imitate the shape of the genuine signature better than even the owner. We apply discriminative features selectively to reject skilled forgeries effectively.
For experiments, we collected 500 genuine signatures, 2,450 random forgeries and 1,000 skilled forgeries. We confirmed that proposed model-based segmentation method gives more stable segmentation results. In an experiments with skilled forgeries, the selective utilization of the features reduced about 62% of error. When
Type I error rate was fixed to 10%, Type II error rate was 0.5% for skilled forgeries and 0% for random forgeries. This result shows the superiority of the model-based segmentation and the selective utilization of the features.