Automatic Signature Verification: In-Depth Investigation of Novel Features and Different Models

摘要

With the growing number of digital pen-input devices used for signature acquisition, forensic handwriting examiners (FHEs) will continue to face a greater number of online signatures in their case work. As a result, the work of scientists in pattern recognition (PR) and the need for automation in the FDEs' casework are beginning to merge. Traditionally, the task of the FDE has been to work with offline signatures that consist of the static trace. The dynamic features of the static trace can only be deduced from the static features, resulting in subjectivity on the part of the FHE. However, with online signatures, additional features such as sequence of strokes, speed, pressure, azimuth and elevation (see list of terms at the conclusion of the paper), can be precisely measured. For the past decade, scientists in PR have been exploring ways to build accurate automated signature systems to see if these systems will provide reliable results when identifying signatures. To begin working more closely with FHEs, a competition of automatic systems on data collected by FHEs has been held in 2009 to test the reliability of automatic signature comparison systems1. As a follow-up, a workshop was initiated in 2011 with both scientists in PR and FHEs discussing their research and daily case work respectively2. The focus in this paper, however, is on the latest experiments in automated signature identification with the 2009 data. These experiments include a novel set of features, the use of Gaussian mixture models (GMMs) and hidden Markov models (HMMs) as classifiers, and several system parameters. In previous experiments, only one reference or comparison signature was used. In the current experiments, several references signatures were used to compare to the questioned signature. Even though this paper focuses on online data, for the purpose of completeness, information about offline data processing will be provided as well.