Humans in the loop: Study of semi-automatic signature recognition based on attributes

Abstract

The present work analyzes performance, abilities and contributions of the human being (layman) in semi-automatic signature recognition systems. During the last decade the performance of Automatic Signature Verification systems have been improved based on new machine learning techniques and better knowledge about intraclass and interclass variability of signers. However, there is still room for improvements and some real world applications demands lower error rates. This work analyzes collaborative tools such as crowdsourcing and human-assisted schemes developed to improve Automatic Signature Verification systems. The performance of humans in semi-automatic recognition tasks is directly related to the information provided during the comparisons. How humans can help automatic systems goes from direct forgery detection to semiautomatic attribute labeling. In this work, we present recent advances, analyzing their performance according to the same experimental protocol. The results suggest the potential of comparative attributes as a way to improve Automatic Signature Verification systems.