Local Features for Forensic Signature Verification

摘要

In this paper we present a novel comparison among three local features based offline systems for forensic signature verification. Forensic signature verification involves various signing behaviors, e.g., disguised signatures, which are generally not considered by Pattern Recognition (PR) researchers. The first system is based on nine local features with Gaussian Mixture Models (GMMs) classification. The second system utilizes a combination of scale-invariant Speeded Up Robust Features (SURF) and Fast Retina Keypoints (FREAK). The third system is based on a combination of Features from Accelerated Segment Test (FAST) and FREAK. All of these systems are evaluated on the dataset of the 4NSigComp2010 signature verification competition which is the first publicly available dataset containing disguised signatures. Results indicate that our local features based systems outperform all the participants of the said competition both in terms of time and equal error rate.