Enhancing camera-captured Devanagari documents via geometric filtering for improved vision-language model text extraction
Anup Kelkar, Parag Deshpande, OG Kakde · 2026
Read Enhancing camera-captured Devanagari documents via geometric filtering for improved vision-language model text extraction by Anup Kelkar, Parag Deshpande, OG Kakde (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/pmc-13226216.txt.
Description
Anup Kelkar, Parag Deshpande, and Kakde improve text reading from photos in a 2026 paper. Vision-language models read English documents well but stumble on Devanagari scripts, common across South Asian languages, especially in crooked hand-held book photos. Their fix cleans the image before the model ever sees it.
Geometric filters tuned to the script's letter shapes, plus clustering, locate Devanagari sentences across font sizes, angles, and worn pages. On book, paper, and pamphlet images the method finds over ninety-five percent of sentences, beating standard cleanup tools. The authors present it as a front step that lifts model accuracy for regional scripts.
Catalog record
| CREATOR | Anup Kelkar, Parag Deshpande, OG Kakde |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |