khmer-ocr-ocm 0.1.0

Khmer OCR (CRNN + CTC) with bundled weights and old Khmer inscription fonts

khmer-ocr-ocm

Khmer OCR built on a CRNN + CTC pipeline, trained primarily on modern Khmer text and fine-tuned on datasets that include old Khmer inscription samples. The package ships with:

• the fine-tuned CRNN weights (model.pth, ~36 MB)
• the matching character vocabulary (vocab.json)
• a set of Khmer fonts — including the Angkor family of old Khmer inscription fonts — used for synthetic dataset generation

Install

pip install khmer-ocr-ocm

The wheel includes the bundled model, so the first prediction works offline without any additional download.

Quick start

from khmer_ocr_ocm import KhmerOCR

ocr = KhmerOCR()                       # auto-picks CUDA if available
text = ocr.predict("line.png")         # OCR one cropped line image
print(text)

Batch a folder of line images to a CSV:

ocr.predict_folder("crops/", out_csv="preds.csv")

Command-line interface:

khmer-ocr-ocm predict line.png
khmer-ocr-ocm predict-folder crops/ --out-csv preds.csv

Old Khmer inscription

The bundled Angkor fonts (ANG-TRAN, Angkor0..Angkor5, PANGKOR) are accessible programmatically:

from khmer_ocr_ocm import get_inscription_font_paths
print(get_inscription_font_paths())

The companion Colab notebook (notebooks/test_pypi_model.ipynb) includes a 5-sample test set that renders inscription-style lines with these fonts and runs them through the model.

Generate synthetic training data

from khmer_ocr_ocm.dataset import generate_dataset

generate_dataset(
    corpus=["ព្រះរាជាណាចក្រកម្ពុជា", "សួស្ដី"],
    out_dir="data/synth",
    use_inscription_fonts=True,        # render with Angkor fonts
)

Continue fine-tuning

from khmer_ocr_ocm.finetune import finetune

finetune(
    train_folder="data/train",
    val_folder="data/val",
    labels_file="labels.txt",
    out_model="crnn_khmer_v22.pth",
    num_epochs=10,
    batch_size=32,
    lr=3e-5,
)

Model architecture

• CNN: 7 conv blocks, BatchNorm, ReLU, spatially-asymmetric MaxPool for width-preserving downsampling
• BiLSTM: 2 layers, hidden = 256, bidirectional
• FC + CTC: 984 classes (983 Khmer/ASCII chars + blank at index 0)
• Input: 1×48×512 grayscale, normalized to [-1, 1]

Files bundled

src/khmer_ocr_ocm/assets/
├── model.pth                 # V21 fine-tuned CRNN (~36 MB)
├── vocab.json                # char2idx / idx2char
└── fonts/
    ├── inscription/          # Angkor* old Khmer fonts
    └── khmer/                # modern Khmer fonts for synthetic data

License

Code: MIT. Fonts retain their original licenses — see LICENSE.