Improved recognition of Telugu handwritten characters through synthetic data generation and attention-based modeling

Abstract

In this study, an enhanced methodology for recognizing handwritten characters in the Telugu language is presented, in which conditional generative adversarial networks (cGANs) and vision transformers (ViTs) are combined to address the inherent challenges of the Telugu script. Challenges such as the complex structure of characters, writer-to-writer variation, and the lack of an annotated handwritten dataset for the Telugu language are addressed using the proposed approach. The ViTs is designed to use self-attention mechanisms to capture spatial dependencies and global contextual information needed for accurate recognition, while cGAN-based data augmentation is used to increase dataset diversity and address class imbalance. The proposed approach outperforms various existing approaches based on CNNs, RNNs, and heuristics, achieving recognition accuracies of 97.89% at the character level and 97.34% at the word level. The results from the ablation study show that both the modules (cGAN augmentation, ViT backbone) contribute independently towards the increase in accuracy. Resolution ablation studies help establish 128 × 128 as the optimal compromise between efficiency and fine strokes. The proposed model uses ViT-B/16, a network architecture with approximately 95 M parameters, 17.6 GFLOPS, and an inference time of around 3 ms per character, and outperforms recent OCR transformer-based models such as Swin-T, DeiT-B, CvT, and TrOCR on similar evaluation datasets. Robust performance in noisy environments and real-world cases is further supported by robustness testing.