SpiS-GAN: Spiral-Modulated Handwriting Synthesis with Star Operation

Handwriting recognition (HTR) systems often struggle with low-resource languages or limited training data. While synthetic handwriting generation can expand these datasets, existing models often fail to capture the fluid, cursive nature of human writing, resulting in broken strokes, blurred edges, or unnatural character shapes. This study introduces SpiS-GAN, a framework designed to generate realistic, writer-consistent handwriting from a single reference image while maintaining computational efficiency.

The authors propose several architectural innovations to address the limitations of traditional CNN-based GANs. The core of the generator is the Star-Spiral Block (SSB), which combines a Modulated Elliptical SpiralFC layer with the star operation. This design allows the model to capture spatial relationships along complex, non-grid trajectories—mimicking the natural flow of a pen—while using element-wise multiplication to expand feature representation without significant computational overhead. To ensure visual fidelity, the authors introduce a Sobel-Regularized Edge Reconstruction Loss (SELoss), which provides explicit guidance to keep stroke boundaries sharp. Finally, the framework employs a dual-discriminator setup: one standard discriminator for overall realism and a second, MLP-based discriminator that evaluates geometric features to detect broken connections or structural anomalies.

SpiS-GAN demonstrates superior performance compared to current state-of-the-art models on both English and Vietnamese handwriting datasets. The generated images are highly authentic and preserve the specific stylistic nuances of the reference writer. Furthermore, the authors show that augmenting training sets with SpiS-GAN-generated data significantly reduces error rates in downstream HTR systems, proving its utility as a practical tool for real-world document analysis and low-resource language processing.