Enhancing Hindi OCR robustness with CRNN-ResNet50: a data augmentation approach on the devanagari dataset.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Optical Character Recognition (OCR) systems face challenges with script variations, font styles, and image quality.
• Existing datasets like IIIT-HW-Dev have deficiencies for robust Hindi OCR.

Purpose of the Study:

• To improve the robustness of OCR models for the Hindi language.
• To address limitations in existing Hindi OCR datasets through data augmentation.

Main Methods:

• Proposed two data augmentation methods: synthetic images (half characters, conjuncts) and image degradations.
• Trained Convolutional Recurrent Neural Networks (CRNN) and ResNet-50 models on the augmented dataset.

Main Results:

• CRNN model achieved a Character Error Rate (CER) of 2.14% and Word Error Rate (WER) of 7.96%.
• ResNet-50 model achieved a CER of 3.27% and WER of 11.82%.
• Demonstrated robustness on complex words and degraded images.

Conclusions:

• Data augmentation significantly enhances Hindi OCR model performance.
• CRNN architecture combined with augmented data provides superior results for Hindi OCR.
• The augmented dataset and models are publicly available to foster further research.