Adapting multilingual vision language transformers for low-resource Urdu optical character recognition (OCR).

Area of Science:

• Natural Language Processing
• Computer Vision
• Digital Humanities

Background:

• Low-resource languages present significant challenges for accurate optical character recognition (OCR).
• Existing OCR systems often struggle with the linguistic complexities of under-resourced languages.
• Digitizing written content in these languages requires specialized approaches.

Purpose of the Study:

• To introduce ViLanOCR, an innovative bilingual OCR system for Urdu and English.
• To address the specific challenges of OCR in low-resource language contexts.
• To demonstrate superior performance compared to existing OCR solutions.

Main Methods:

• Development of ViLanOCR, a bilingual OCR system.
• Leveraging advanced multilingual transformer-based language models.
• Evaluation using the character error rate (CER) metric on the Urdu UHWR dataset.

Main Results:

• ViLanOCR achieves a character error rate (CER) of 1.1% on the Urdu UHWR dataset.
• The system demonstrates state-of-the-art performance for Urdu handwriting digitization.
• Experimental results confirm the effectiveness of the proposed approach.

Conclusions:

• ViLanOCR offers a robust solution for OCR in low-resource languages like Urdu.
• Advanced transformer models are effective for improving OCR accuracy in challenging linguistic contexts.
• The system surpasses current state-of-the-art baselines in Urdu handwriting digitization.