AKSHARA VISION: A SWIN TRANSFORMER– BASED FRAMEWORK FOR ROBUST KANNADA HANDWRITTEN CHARACTER RECOGNITION WITH PROGRESSIVE SEGMENTATION AND HUMAN-IN-THE-LOOP FEEDBACK

Abstract

Handwritten character recognition for Indic scripts presents fundamental challenges due to dense ligatures, diacritic interactions and high intra-class variation, combined with the absence of standardized large-scale datasets. Kannada script adds further complexity through its many conjunct forms and visually similar glyphs. This work introduces AksharaVision, a complete handwritten OCR framework grounded in the Swin Transformer architecture. The system integrates a curated 113-class Kannada dataset, a Swin-Tiny backbone with centroidbased embedding refinement, a progressive segmentation strategy enabling prototype word recognition without sequence-level training, and a feedback-enabled Gradio interface. Evaluations across balanced augmentation, focal-loss optimisation, cosine learning schedules, and calibration demonstrate strong generalization, achieving 99.89% Top-1 accuracy and an Expected Calibration Error (ECE) of 0.001. Diagnostic analyses including confusion studies, robustness testing and embedding behaviour further highlight the reliability of the proposed model, establishing a foundation for scalable OCR in low-resource Indic scripts.