[Paper Review] Tamizhi-Net OCR: Creating A Quality Large Scale Tamil-Sinhala-English Parallel Corpus Using Deep Learning Based Printed Character Recognition (PCR).

Charangan Vasantharajan, Uthayasanker Thayasivam|arXiv (Cornell University)|Sep 13, 2021

Handwritten Text Recognition Techniques9 references4 citations

TL;DR

This paper proposes Tamizhi-Net OCR, a deep learning-enhanced Tesseract 4.1.1 model using LSTM-based training on legacy fonts to improve printed character recognition (PCR) in Tamil, Sinhala, and English. It achieves 98.2% character-level accuracy for Tamil (+12.9% relative improvement) and 94.8% for Sinhala (+3.26% relative improvement), enabling scalable, automatic extraction of multilingual text from PDFs with non-standard fonts.

ABSTRACT

Most of the low resource languages do not have the necessary resources to create even a substantial monolingual corpus. These languages may often be found in government proceedings but mostly in the form of Portable Document Formats (PDFs) that contains legacy fonts. Extracting text from these documents to create a monolingual corpus is challenging due to legacy font usage and printer-friendly encoding which are not optimized for text extraction. Therefore, we propose a simple, automatic, and novel idea that can scale for Tamil, Sinhala, and English languages and many documents. For this purpose, we enhanced the performance of Tesseract 4.1.1 by employing LSTM-based training on many legacy fonts to recognize printed characters in the above languages. Especially, our model detects code-mix text, numbers, and special characters from the printed document. It is shown that this approach can boost the character-level accuracy of Tesseract 4.1.1 from 85.5 to 98.2 for Tamil (+12.9% relative change) and 91.8 to 94.8 for Sinhala (+3.26% relative change) on a dataset that is considered as challenging by its authors.

Motivation & Objective

• To address the lack of large-scale, high-quality monolingual corpora for low-resource languages like Tamil and Sinhala.
• To overcome challenges in extracting text from PDFs with legacy fonts and printer-friendly encodings.
• To develop an automatic, scalable method for recognizing printed characters in Tamil, Sinhala, and English in mixed-language (code-mixed) documents.
• To improve the character-level accuracy of Tesseract 4.1.1 for these languages using deep learning on legacy font data.
• To enable the creation of a parallel Tamil-Sinhala-English corpus from scanned, legacy-font PDFs.

Proposed method

• Enhanced Tesseract 4.1.1 with LSTM-based optical character recognition (OCR) models trained on diverse legacy fonts.
• Trained the model on a large-scale, multilingual dataset containing printed Tamil, Sinhala, and English characters, including code-mixed text, numbers, and special characters.
• Applied transfer learning techniques to adapt Tesseract’s existing architecture to recognize non-Latin scripts with complex character sets.
• Optimized the model for handling printer-friendly encodings and non-standard font mappings commonly found in government documents.
• Used a custom data pipeline to preprocess and augment legacy-font PDFs into training samples for improved generalization.
• Evaluated performance using a challenging benchmark dataset to measure character-level accuracy improvements.

Experimental results

Research questions

• RQ1 Can LSTM-enhanced Tesseract 4.1.1 significantly improve character-level accuracy for Tamil and Sinhala in documents with legacy fonts?
• RQ2 To what extent can the model detect and correctly recognize code-mixed text, numbers, and special characters in multilingual printed documents?
• RQ3 How scalable is the proposed method for processing large volumes of government and archival documents in low-resource languages?
• RQ4 What is the relative improvement in OCR accuracy for Tamil and Sinhala compared to baseline Tesseract 4.1.1 on legacy font PDFs?
• RQ5 Can this approach enable the automatic creation of a high-quality, parallel Tamil-Sinhala-English corpus from scanned documents?

Key findings

• The proposed Tamizhi-Net OCR model improved character-level accuracy for Tamil from 85.5% to 98.2% on a challenging benchmark dataset, representing a 12.9% relative improvement.
• For Sinhala, the model achieved 94.8% character-level accuracy, a 3.26% relative improvement over the baseline Tesseract 4.1.1's 91.8%.
• The model successfully recognized code-mixed text, numbers, and special characters with high fidelity in multilingual printed documents.
• The approach demonstrated scalability for processing large volumes of legacy-font PDFs commonly found in government and archival repositories.
• The enhanced Tesseract model enables the automatic creation of high-quality monolingual and parallel corpora from scanned, non-standard document formats.
• The results confirm that LSTM-based fine-tuning on legacy fonts significantly enhances OCR performance for low-resource scripts like Tamil and Sinhala.