Writer recognition enhancement by means of synthetically generated handwritten text

This paper presents a new method to generate synthetic executions of on-line words from real samples. The proposed generation method takes advantage of the characteristics of a writer recognition system developed by the authors and can be seamlessly integrated into it. Both the generation method and the recognition system consider strokes as the structural units of handwriting with words being regarded as two sequences, one of pen-up and one of pen-down strokes. Given two samples from the same word and writer, a new sample is produced by aligning their sequences of strokes and then averaging the matching pairs. Thanks to a self-organising map used to categorise strokes, the alignment and comparison of sequences of strokes are performed in a straightforward and computationally efficient way.The synthetically generated words not only retain much of the discriminative power (i.e. the capability to discriminate among writers) of the words from which they are generated, but in some cases exhibit an increased recognition performance. Also, the newly generated words allow enlarging the number of available samples in the enrolment sets that are used to build writers' models. In most cases, this enlargement has the effect to improve the performance of the recognition system. Experimenting with 320 writers and enrolment sets containing 3 real samples and 6 synthetically generated ones, verification is improved for 15 of the 16 words in the BiosecurID database, with the minimum of the detection cost function being reduced by up to a 26.5%.