Robust multilingual OCR: from Ancient Indic texts to modern India Street signs

Abstract

Indic Texts often contain out-of-vocabulary (OOV) words leading to significant errors in text scanned with Optical Character Recognition (OCR). We present an interactive framework for adaptively assisting Indic OCR corrections. Subsequently, we propose OCR correction methods for resource-constrained settings, such as multi-OCR consensus, plug-in classifiers, and LSTMs with a fixed delay.\n\nModern Indian street signs and license plates pose an even tougher reading challenge. They often appear in a variety of languages, fonts, sizes, and orientations. We present the first scene text recognition results using multi-headed attention models.