A Writer-Dependent Approach to Off-line Signature Verification

Abstract

Results of a new approach to off-line signature verification are presented. The approach is writer-dependent. To verify a signature, only 15 ≥N ≥5 genuine signatures of the person are used. The signature images are pre-processed and normalized into a contour representation. We then compute two new signature features: the distribution of LBP values and local curvature of contours in the binary signature image. For a signature submitted for analysis, N genuine signatures of this person are randomly selected and a one-class SVM classifier is developed. Accuracy of our approach in verification of all 2640 signatures from the public CEDAR database was 99.77%. All fake signatures were correctly recognized even with N=5 genuine signatures used to build the classifier