Personalizing Handwriting Recognition Systems with Limited User-Specific Samples

Christian Gold, Dario van den Boom, Torsten Zesch

Abstract: Personalization of handwriting recognition is still an understudied area due to the lack of a comprehensive dataset. We collect a dataset of 37,000 words handwritten by 40 writers that we make publicly available. We investigate the impact of personalization on recognition by training a baseline recognition model and retraining it using our dataset. After controlling that our model really adapts to the personal handwriting style and not just to the overall domain, we show that personalization in general requires several hundred samples to be effective. However, we show that the choice of transfer samples is important and that we can quickly personalize a model with a limited number of samples. We also examine whether we can detect adversarial behavior trying to reduce recognition performance.