HTLDIFF: Handwritten Text Line Generation - A Diffusion Model Perspective
Aniket Gurav, Sukalpa Chanda, Narayanan Chatapuram Krishnan · 2024
The ability to generate realistic synthetic text lines is crucial for training robust handwritten text recognition models when there is a scarcity of annotated training data. Existing generative models for handwritten text generation have limitations in handling the variable length nature of lines and capturing spatial relationships between words. We present a novel Diffusion-based method HTLDIFF that enables the generation of stylized handwritten text lines. Our approach demonstrates the ability to produce high-quality synthetic lines of handwritten text across diverse writing styles. By training text recognition systems on a combination of original and generated data, we observed enhanced performance of handwritten text recognition systems. Moreover, when utilizing the original training data for a writer identification system, we achieved good accuracy in recognizing the unique styles exhibited in the generated writing samples. This demonstrates that our generated texts effectively emulate styles, enhancing both content and writer identification capabilities. The proposed diffusion-based approach effectively addresses challenges in handwritten line generation, resulting in the generation of stylistically consistent synthetic data.