SemiHastakshar
Semi-supervised Indic handwritten OCR, trained on large-scale unlabeled handwriting collected from the web.
Problem. HTR models trained only on small, curated benchmarks struggle to generalize to the messy variety of real-world Indic handwriting — writing styles, scripts, and image quality vary far more “in the wild” than in any lab dataset.
Approach. We collect Indic-HW-Wild, a 2.4M+-word unlabeled corpus of real-world Indic handwriting scraped from the internet, and train with high-confidence pseudo-labeling — iteratively re-labeling the unlabeled pool with the current best model and retraining on the mix.
Results. SemiHastakshar cuts out-of-domain WER by up to 19% (Telugu) over a PARSeq HTR baseline trained only on labeled data, published at ICVGIP 2025 (Evani et al., 2025).
Links. Project page · Paper · Code
