Java-based Tibetan translation and text-processing application focused on
Buddhist terminology workflows. It combines dictionary lookup,
tokenization, transliteration, ranking/model services,
and optional OCR-related utilities.
Deep learning research project comparing lightweight CNN and
CNN+Transformer hybrid architectures for Tibetan character OCR.
Focused on replacing hand-crafted OCR pipelines with learned
convolutional and attention-based feature extraction.
Delivered as namsel_buda — a full modernization of the
classic Namsel OCR engine
(Python 3.12, the CNN recognizer above, and safe pickle-free model
formats). It is the Tibetan OCR engine embedded in TradutorBUDA,
turning scanned pages into editable Unicode for dictionary lookup and
translation.
Tech:
PyTorch, CNNs, Transformers, ONNX, INT8 Quantization,
Gradio, Hugging Face Spaces, Mixed Precision Training
95.52% validation accuracy on 1,020 Tibetan character classes
Lightweight CNN and CNN+Transformer hybrid comparison
Interactive deployment using Hugging Face Spaces
Shipped as namsel_buda — a modernized fork of the
Namsel OCR engine (Python 3.12, CNN recognizer, pickle-free
data-only model formats)