CRB-GAN: Generating Diverse Samples for Single-sample Character Classes in Handwritten Historical Documents

Enhancing sample diversity of minority classes within an extremely imbalanced handwritten character dataset poses a formidable challenge for GAN-based frameworks. This paper proposes Character Rebalance GAN (CRB-GAN), a novel deep generative model designed to address extreme class imbalance in historical handwritten recognition, especially for scripts like Chinese and Tibetan with many single-sample characters. Unlike prior approaches, CRB-GAN introduces a latent variable initialization method that aligns the mean and covariance of machine-printed and handwritten character embeddings, enabling diversity even before adversarial training. It further incorporates a class-aware loss strategy to balance fidelity and diversity. Experimental result shows that our method not only surpasses others in terms of sample quality and diversity but also demonstrates notable improvements in classification accuracy. Specifically, accuracy sees an impressive rise from 89.26% to 93.83% for Tibetan, and 95.36% to 97.53% for Chinese. CRB-GAN requires no language-specific morphological features, offering a universal, scalable solution for augmenting rare handwritten character classes.