Discriminative power of online handwritten words for writer recognition

This paper is aimed at exploring the potential of online words to perform biometric writer recognition. Most of the scientific literature dealing with online writer recognition has focused on signature and somehow disregarded handwritten text. Using a novel recognition system based on stroke categorization and dynamic time warping, it is shown that short sequences of online text (words and combinations of a small number of words) perform remarkably well not only in identification but also in verification. Experimentation is performed with a dataset comprising 320 writers who donated 4 repetitions of 16 different Spanish words. With a single word, identification rate ranges from 80.6% to 95.6% and verification error from 1.57% to 5.55%. When two words are combined, the best identification rate increases up to 99.7% and the best verification error decreases until 0.63%.