Indic Handwritten Text Recognition
Transformer-based OCR (PARSeq) for handwritten text across ten Indic scripts, boosted by transfer learning from printed text.
Problem. Handwritten Text Recognition (HTR) is hard even for a single script — writing styles, noise, and imperfections vary widely. It gets harder across Indic scripts, which have complex character structures, large character inventories, and comparatively little labeled data.
Approach. We adapt PARSeq — a transformer model originally built for scene-text recognition — to handwritten text across ten Indic scripts (Hindi, Bengali, Telugu, Tamil, Gujarati, Gurumukhi, Odia, Kannada, Malayalam, Urdu), and study how much pre-training on printed text before fine-tuning on handwritten text helps.
Results. PARSeq beats prior CNN–RNN baselines on most of the ten languages, and transfer learning from printed to handwritten text gives a further, consistent accuracy boost — published at CVIP 2024 (Evani et al., 2024).
Links. Project page · Paper · Code
