[Paper] Handwriting Imitation without Real Training Samples based on Augmented Digital Font Text Images and Test-time Fine-tuning

Abstract

Handwriting imitation, a task of generating text images that imitate a target writer's handwriting style, has been widely studied. However, existing methods require a massive amount of real handwriting samples to train a handwriting imitation model, which are impractical to collect. This paper proposes a novel approach that eliminates the need for real handwritten texts as training samples. Instead, we train a handwriting imitation model on digital font text images (DFTI), which are much easier to obtain. To address the limited stylistic variation of DFTI, we introduce a novel Elastic Transform-based data augmentation technique that ensures style consistency across all characters in a single text image. Furthermore, to strengthen the imitation performance, we apply test-time fine-tuning to the trained model, utilizing a target writer's style reference images. Experiments on ten writers selected from the IAM-OnDB dataset demonstrate that our method achieves competitive or superior imitation quality compared to a baseline model trained on a large set of real handwriting samples.