Script-agnostic Reflow of Text in Document Images

Reading text from document images can be difficult on mobile devices due to the limited screen width available on them. While there exist solutions for reflowing Latin-script texts on such devices, these solutions do not work well for images of other scripts or combinations of scripts, since they rely on script-specific characteristics or OCR. We present a technique that reflows text in document images in a manner that is agnostic to the script used to compose them. Our technique achieved over 95% segmentation accuracy for a corpus of 139 images containing text in 4 genetically-distant languages-English, Hindi, Kannada and Arabic. A preliminary user study with a prototype implementation of the technique provided evidence of some of its usability benefits.