Khmer Handwritten OCR
Reads handwritten Khmer from full pages.
- Role
- Researcher
- Context
- B.Sc. thesis · RUPP
- Stack
- PyTorch · Python · Hugging Face

Most OCR tools are built for printed Latin text and struggle with handwritten Khmer — stacked consonants, vowel signs above and below the line, and no spaces between words. My final-year thesis builds a complete system that takes a photo of a handwritten page, finds every line of text and turns it into editable Khmer Unicode.
What I did
- 01
Built a text-line detector that finds every line on a page photo — even tilted, shadowed or on lined paper.
- 02
Trained a line recognizer (CRNN with CTC) that converts each line image into Khmer text.
- 03
Generated a synthetic handwriting dataset to pre-train the recognizer before fine-tuning on real handwriting.
- 04
Combined both models into a full-page pipeline with a demo app for uploading photos.
- 05
Published the models openly on Hugging Face for other researchers and developers.
Highlights
Full-page reading
Line detection and recognition in one pipeline.
Synthetic data
Generated handwriting to learn rare characters.
Lightweight
Small models that run on a regular CPU.
Open model
Released on Hugging Face.
Gallery

