Spatio-Temporal Handwriting Imitation

  • 2021-04-16 17:09:48
  • Martin Mayr, Martin Stumpf, Anguelos Nicolaou, Mathias Seuret, Andreas Maier, Vincent Christlein
  • 1

Abstract

Most people think that their handwriting is unique and cannot be imitated bymachines, especially not using completely new content. Current cursivehandwriting synthesis is visually limited or needs user interaction. We showthat subdividing the process into smaller subtasks makes it possible to imitatesomeone's handwriting with a high chance to be visually indistinguishable forhumans. Therefore, a given handwritten sample will be used as the target style.This sample is transferred to an online sequence. Then, a method for onlinehandwriting synthesis is used to produce a new realistic-looking text primedwith the online input sequence. This new text is then rendered andstyle-adapted to the input pen. We show the effectiveness of the pipeline bygenerating in- and out-of-vocabulary handwritten samples that are validated ina comprehensive user study. Additionally, we show that also a typical writeridentification system can partially be fooled by the created fake handwritings.

 

Quick Read (beta)

loading the full paper ...