A Novel Machine Annotated Balanced Bangla OCR Corpus

Abstract

We present a balanced and 100% machine annotated Bangla OCR corpus of nearly eight and a half million characters. This is a "first-of-its-kind" effort for the Bangla language. Although Bangla is a top five most frequently used language in the world, it is still considered to be a resource-poor language since even the most basic language processing tools are not available. This corpus goes a long way in mitigating this shortcoming. Also, it is important to mention that this is a continuation of our previous work in building a synthetic corpus for training Bangla OCRs and is an intermediate step in our effort to build a true gold-standard corpus for Bangla OCRs-our ultimate goal. The paper not only discusses the corpus, but also the entire processing pipeline-from scanning pages to identifying lines to segmenting words and finally characters-which are then annotated using a homegrown OCR. We also present a comprehensive corpus characteristics and specification review-demonstrating that we have built a corpus which is very nearly balanced-a very desirable characteristic of any corpus design.