ET-Network: A novel efficient transformer deep learning model for automated Urdu handwritten text recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Urdu handwritten text recognition (UHTR) is challenging due to inconsistent writing styles and limited data, unlike high-resource languages.
• Existing optical character recognition (OCR) methods struggle with the complexities of Urdu's cursive script and data scarcity.
• Transformer models show promise for UHTR, addressing limitations of traditional approaches.

Purpose of the Study:

• To develop an advanced transformer-based method for accurate Urdu handwritten text recognition.
• To enhance feature extraction by integrating self-attention mechanisms into EfficientNet.
• To establish a new state-of-the-art performance benchmark for UHTR.

Main Methods:

• Proposed ET-Network, combining EfficientNet with self-attention for robust feature extraction.
• Utilized a vanilla transformer architecture for language modeling and text generation.
• Employed prefix beam search for optimizing recognition outcomes.
• Trained and evaluated the model on three diverse Urdu handwritten datasets: NUST-UHWR, UPTI2.0, and MMU-OCR-21.

Main Results:

• Achieved a 4% reduction in character error rate (CER) and a 1.55% reduction in word error rate (WER).
• Established a new state-of-the-art CER of 5.27% and WER of 19.09% for Urdu handwritten text recognition.
• Demonstrated the effectiveness of integrating self-attention for capturing long-range dependencies in Urdu script.

Conclusions:

• The ET-Network offers a significant improvement over existing methods for Urdu handwritten text recognition.
• Transformer-based approaches, particularly with self-attention, are highly effective for low-resource language OCR.
• The established state-of-the-art performance highlights the potential of ET-Network for practical applications.