Devanagari Text Recognition: A Transcription Based Formulation

Abstract

Abstract-Optical Character Recognition (OCR) problems are often formulated as isolated character (symbol) classification task followed by a post-classification stage (which contains modules like Unicode generation, error correction etc. ) to generate the textual representation, for most of the Indian scripts. Such approaches are prone to failures due to (i) difficulties in designing reliable word-to-symbol segmentation module that can robustly work in presence of degraded (cut/fused) images and (ii) converting the outputs of the classifiers to a valid sequence of Unicodes. In this paper, we propose a formulation, where the expectations on these two modules is minimized, and the harder recognition task is modelled as learning of an appropriate sequence to sequence translation scheme. We thus formulate the recognition as a direct transcription problem. Given many examples of feature sequences and their corresponding Unicode representations, our objective is to learn a mapping which can convert a word directly into a Unicode sequence. This formulation has multiple practical advantages: (i) This reduces the number of classes significantly for the Indian scripts. (ii) It removes the need for a reliable wordto-symbol segmentation. (ii) It does not require strong annotation of symbols to design the classifiers, and (iii) It directly generates a valid sequence of Unicodes. We test our method on more than 6000 pages of printed Devanagari documents from multiple sources. Our method consistently outperforms other state of the art implementations.

Towards a Robust OCR System for Indic Scripts

“…The system should be capable of learning these rules. While [7] demonstrated the system which recognizes Hindi (a language with relatively fewer rearranging rules), the present work shows the result on multiple Indic scripts, having much complicated rearranging rules. To accommodate such rules it is important for the classifier to analyze the feature segment based on its forward and backward information.…”

“…More over, these conversions are often brittle and can fail with noisy symbol labels. In our previous work [7], we have proposed a solution to this problem by considering it as a sequence to sequence transcription, where we convert the sequence of word features into the corresponding text sequence. Such an approach does not mandate one to identify the UNICODE rearranging rules in advance.…”

“…Advances in machine learning has also enabled the output prediction as structured prediction task rather than predicting individual class labels and then combining them to obtain semantically richer structures such as a list of class labels or a list of UNICODEs. (Note that for Indic scripts often the classes are defined very differently from the codes used for the representation.) This paper builds on some of our recent attempts [3], [7], [10] to design a generic framework, which is also applicable for multiple Indian scripts.…”

“…Most of the previous attempts have been script specific and required high amount of expertise in transferring the knowledge across the scripts. Extensive use of machine learning (not just the design of classifiers for isolated character segments) have helped the document image analysis in multiple ways in the recent years [2], [4], [5], [6], [7], [8]. For scripts/languages, where the research manpower (and also possibly the commercial interests) are lacking, this has greatly helped in rapidly building reasonably robust systems.…”

Text recognition using deep BLSTM networks

“…Some work has also been done in the last few years on online handwritten recognition of Telugu script using HMM [23] and online handwritten Tamil word recognition [24] have used segmentation based approaches. Naveen et al presented a direct implementation of single layer LSTM network for the recognition of Devanagiri scripts [25], [26] and further experimented on more Indic scripts [27]. In this paper we present a Deep BLSTM based RNN architecture that is script independent, segmentation free and does not have any unicode re-ordering issues.…”