Sanskrit OCR and Translation of degraded Sanskrit manuscripts using Deep Learning
Check out the live demo and results here:
🔗 LinkedIn Demo Post
Due to computational and GPU hardware limitations, all training, inference testing, and integration steps were performed on a single local machine. This constraint required centralized development, and thus all Git operations (including commits and pushes) originated from the same device.
Advantages of this approach:
- Ensured consistent CUDA and PyTorch environment for model training
- Avoided serialization/deserialization issues with large-scale checkpoints
- Direct and quick access to logs, weights, and intermediate outputs
- Object Detection Backbone: Ultralytics YOLOv8
- OCR Engine: Microsoft TrOCR (Transformer OCR)
- Auxiliary OCR: Google Cloud Vision API
- Image Preprocessing & Feature Extraction: Torch + TorchVision
- mAP@0.5:
0.995(Mean Average Precision at IoU threshold 0.5) - F1 Confidence Curve Score:
1.00 - Confusion Matrix: Flawless performance on actual Sanskrit manuscript datasets.
➕ Refer to theresults/folder for sample predictions and benchmarking screenshots.
- Ultralytics YOLOv8
- Torch + TorchVision
- TrOCR
- Google Cloud Vision API
- TypeScript
- JavaScript
- Vite
- Tailwind CSS
- Python
- FastAPI
| Layer | Repository Link |
|---|---|
| Frontend | GitHub: OBZIUS |
| Backend | GitHub: NeuralSynth |
| Model Training | GitHub: NoiceHax |
- Sample predictions, bounding box overlays, and post-OCR outputs can be found in the
results/directory.
- Migrate training to cloud-based GPUs for higher scalability
- Add multilingual support and improve layout analysis
- Build a full user-facing dashboard for batch PDF/document OCR