Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Deployable OCR models for Indic languages

Minesh Mathew, Ajoy Mondal, CV Jawahar

TL;DR

The paper addresses OCR for Indian languages with diverse scripts by conducting a comprehensive empirical study of CTC-based transcription across 13 official languages using a newly introduced Mozhi dataset. It systematically compares encoder architectures (Col_RNN, Win_RNN, CNN_only, CRNN), demonstrates that CRNN generally yields the best word/line recognition performance, and shows that a pipeline combining segmentation tools with CRNN-based recognition achieves strong end-to-end page OCR results that surpass public tools on several languages. In addition to reporting recognition performance, the authors release Mozhi (1.2M word annotations across 13 languages) and provide publicly accessible APIs and a web app to facilitate digitization of Indic documents, aiming to accelerate research and deployment. The work highlights practical applicability, reveals current limitations in script/diversity and font variations, and outlines directions for broader data collection and pretraining to improve robustness across diverse documents.

Abstract

Recognition of text on word or line images, without the need for sub-word segmentation has become the mainstream of research and development of text recognition for Indian languages. Modelling unsegmented sequences using Connectionist Temporal Classification (CTC) is the most commonly used approach for segmentation-free OCR. In this work we present a comprehensive empirical study of various neural network models that uses CTC for transcribing step-wise predictions in the neural network output to a Unicode sequence. The study is conducted for 13 Indian languages, using an internal dataset that has around 1000 pages per language. We study the choice of line vs word as the recognition unit, and use of synthetic data to train the models. We compare our models with popular publicly available OCR tools for end-to-end document image recognition. Our end-to-end pipeline that employ our recognition models and existing text segmentation tools outperform these public OCR tools for 8 out of the 13 languages. We also introduce a new public dataset called Mozhi for word and line recognition in Indian language. The dataset contains more than 1.2 million annotated word images (120 thousand text lines) across 13 Indian languages. Our code, trained models and the Mozhi dataset will be made available at http://cvit.iiit.ac.in/research/projects/cvit-projects/

Towards Deployable OCR models for Indic languages

TL;DR

The paper addresses OCR for Indian languages with diverse scripts by conducting a comprehensive empirical study of CTC-based transcription across 13 official languages using a newly introduced Mozhi dataset. It systematically compares encoder architectures (Col_RNN, Win_RNN, CNN_only, CRNN), demonstrates that CRNN generally yields the best word/line recognition performance, and shows that a pipeline combining segmentation tools with CRNN-based recognition achieves strong end-to-end page OCR results that surpass public tools on several languages. In addition to reporting recognition performance, the authors release Mozhi (1.2M word annotations across 13 languages) and provide publicly accessible APIs and a web app to facilitate digitization of Indic documents, aiming to accelerate research and deployment. The work highlights practical applicability, reveals current limitations in script/diversity and font variations, and outlines directions for broader data collection and pretraining to improve robustness across diverse documents.

Abstract

Recognition of text on word or line images, without the need for sub-word segmentation has become the mainstream of research and development of text recognition for Indian languages. Modelling unsegmented sequences using Connectionist Temporal Classification (CTC) is the most commonly used approach for segmentation-free OCR. In this work we present a comprehensive empirical study of various neural network models that uses CTC for transcribing step-wise predictions in the neural network output to a Unicode sequence. The study is conducted for 13 Indian languages, using an internal dataset that has around 1000 pages per language. We study the choice of line vs word as the recognition unit, and use of synthetic data to train the models. We compare our models with popular publicly available OCR tools for end-to-end document image recognition. Our end-to-end pipeline that employ our recognition models and existing text segmentation tools outperform these public OCR tools for 8 out of the 13 languages. We also introduce a new public dataset called Mozhi for word and line recognition in Indian language. The dataset contains more than 1.2 million annotated word images (120 thousand text lines) across 13 Indian languages. Our code, trained models and the Mozhi dataset will be made available at http://cvit.iiit.ac.in/research/projects/cvit-projects/
Paper Structure (20 sections, 7 equations, 7 figures, 4 tables)

This paper contains 20 sections, 7 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Mozhi Dataset
  4. APIs and Web-based Applications
  5. Text Recognition using CTC Transcription
  6. Extracting Feature Sequence
  7. Encoder:
  8. Decoder:
  9. Training
  10. Inference
  11. Experimental Setup
  12. Implementation Details
  13. Evaluation
  14. Experiments and Results
  15. Comparing Different Encoder Configurations
  16. ...and 5 more sections

Figures (7)

  • Figure 1: We explore printed text recognition across 13 Indian languages, covering ten unique scripts. Although many languages share a common alphabet, their scripts vary, with exceptions like Hindi and Marathi. The last column shows the name "Gandhi" in all ten scripts.
  • Figure 2: Shows a few sample of cropped images of each of 13 languages from our Mozhi dataset.
  • Figure 3: A few sample of word level images from our Mozhi dataset.
  • Figure 4: Shows screen shot of our web-based APIs to digitize Indic printed documents.
  • Figure 5: We examine four CTC-based text recognition methods --- Col_RNN, Win_RNN, CNN_only, and CRNN, distinguished by their feature extraction and sequence encoding. $W$ and $H$ represent the width and height of the input image $I$, respectively. $|L^\prime|$ indicates the number of class labels, including the blank label. $Hid_j$ signifies the number of hidden units in the last RNN layer. In the case of Win_RNN, $W_W$, and $S_W$ denote the width and step size of the sliding window, respectively.
  • ...and 2 more figures