Systematic evaluation of pre-processing approaches in online signature verification

Article type: Research Article

Authors: Saleem, Mohammad^* | Lia Szucs, Cintia | Kovari, Bence

Affiliations: Budapest University of Technology and Economics, Hungary

Correspondence: [*] Corresponding author: Mohammad Saleem, Budapest University of Technology and Economics, Hungary. E-mail: msaleem@aut.bme.hu.

Abstract: Online signature verification is a currently evolving field. Numerous verification approaches are proposed each year that are claimed to be capable of delivering improved results in some aspects over others. These systems are usually based on the experiences of the author(s), and the results are often limited to single databases. These factors make it difficult to compare or even reproduce the results. The lack of negative results and evaluation of design choices make the creation of new, improved online verification systems difficult. In this work, we addressed this problem by conducting a systematic evaluation of the most common pre-processing techniques used in conjunction with dynamic time warping-based classification. We evaluated six different pre-processing methods on five databases and devoted over 30 000 computing hours to test 211 680 different verifier configurations. We identified several dead ends, narrowed down the design choices to the most promising configurations, and introduced a simple yet competitive configuration. To aid the research community, we have made all our results, data, and source code public. Our results may have broad implications for the signature verification field as they provide definitive answers for some questions that were previously addressed only through intuition or limited experimentation.

Keywords: Classifier design and evaluation, handwriting analysis, similarity measures, feature evaluation and selection

DOI: 10.3233/IDT-220247

Journal: Intelligent Decision Technologies, vol. 17, no. 3, pp. 655-672, 2023