Towards a standard Bangla PhotoOCR: Text detection and localization

Abstract

A complete Bangla PhotoOCR requires a series of carefully chosen algorithms. Text extraction from images is a long-standing active research area. It is even more attractive today due to the availability of low-cost mobile image acquisition devices. Many researchers have addressed this problem using different approaches. Often times, the first step towards text extraction from images is detection of text areas. Bangla texts, specially in images, pose a unique set of challenges than texts in other languages. In this paper, we experiment with two established approaches, available for other languages, to automatically localize Bangla texts in complex natural scene images towards developing a complete Bangla PhotoOCR system. In our approach, features are extracted from an image using wavelets based decomposition and histogram calculation techniques. We use 56 features to train two different types of classifiers (ANN based and SVM based) to localize Bangla texts in natural scene images. Our experimental results show that ANN is a good classifier for identifying Bangla texts.