TL;DR
Researchers have developed a high-accuracy OCR system for the ancient Ashokan Brahmi script by leveraging transfer learning with Convolutional Neural Networks (CNNs). By fine-tuning a MobileNet architecture on a custom-augmented dataset, the team achieved a validation accuracy of nearly 96%, offering a powerful new tool for epigraphists to digitize and preserve India’s linguistic heritage.
What Problem Does This Paper Address?
Ashokan Brahmi is a foundational script that emerged in the 3rd century BCE, serving as the ancestor to many modern Indic scripts. However, the ability to read it was lost for centuries until its rediscovery in the 1800s. Today, the study of these ancient inscriptions (epigraphy) is a bottlenecked process: manual transcription is painstaking, and most modern OCR technologies are designed for popular, contemporary languages.
The primary challenge in creating OCR for historical scripts is the "data desert"—there simply aren't enough high-quality, labeled digital samples to train deep learning models from scratch. This paper seeks to bridge that gap by applying modern computer vision techniques to automate the recognition of these ancient characters.
Methods
The researchers employed a multi-stage pipeline to transform raw inscription images into machine-readable text:
- Image Preprocessing: To handle the "noisy" nature of stone inscriptions, they used Median Blur to eliminate salt-and-pepper noise and Otsu’s Thresholding to cleanly separate the foreground characters from the background.
- Segmentation: They utilized horizontal and vertical projection profiles to isolate individual lines and then individual characters from a block of text.
- Transfer Learning: Instead of training a model from zero, they took three established architectures—LeNet, VGG-16, and MobileNet—pre-trained on massive datasets and "fine-tuned" them for Brahmi.
- UI Development: Beyond the model, the team built a suite of tools including Abhiñāna (the OCR tool), Akṣarāntara (for transliteration), and Kośa (a database repository).
Datasets
The team sourced their initial data from Indoskript, a paleographic database of Brahmi and Kharosthi scripts.
- Initial Set: 3,500 high-resolution, low-noise images covering 214 unique characters (6 vowels and 208 consonants).
- Data Augmentation: To overcome the small sample size, they applied random rotations, scaling, shifting, and brightness adjustments.
- Final Set: This process expanded the dataset to approximately 227,000 images (roughly 1,000 variations per character), providing the volume necessary for robust deep learning.
Key Insights and Ideas
- Efficiency Wins: Surprisingly, the lightweight MobileNet outperformed the much deeper VGG-16. This suggests that for character recognition tasks where spatial hierarchies are relatively consistent, the depthwise separable convolutions in MobileNet are more than sufficient and perhaps less prone to the overfitting seen in larger models.
- Pooling Strategy Matters: Within the MobileNet experiments, Average Pooling significantly outperformed Max Pooling (95.94% vs 92.43%). This indicates that for ancient scripts, where stroke thickness and edges might vary due to weathering, capturing the average feature presence is more effective than just picking the strongest signal.
- The Power of Transfer: The study reinforces that transfer learning is the "silver bullet" for low-resource historical languages, allowing models to use general visual knowledge (edges, curves) to learn specific ancient glyphs rapidly.
Key Results
The researchers compared three models, with MobileNet emerging as the clear winner:
| Model | Validation Accuracy | Validation Loss |
|---|---|---|
| MobileNet (Average Pooling) | 95.94% | 0.129 |
| MobileNet (Max Pooling) | 92.43% | 0.2507 |
| LeNet | 70.89% | 1.0673 |
| VGG-16 | 65.82% | 1.2654 |
The high accuracy and low loss of the MobileNet model make it suitable for real-time inference on mobile devices, which is critical for researchers working in the field at archaeological sites.
Limitations and Future Work
While the results are impressive, the authors note several areas for growth:
- Script Variety: The current model is specialized for Ashokan Brahmi. Future iterations aim to expand the dataset to include a broader range of ancient scripts and regional variations.
- Architectural Upgrades: The team plans to explore hybrid models and attention mechanisms, which could help the model focus on specific distinctive strokes of complex characters.
- Real-world Robustness: While the dataset used "less-noise" images, expanding the system to handle highly weathered or damaged inscriptions found in situ remains a significant challenge for the field.
Summary updated