pndaza/burmese-kraken-ocr

Burmese Kraken OCR

A Kraken recognition model for Burmese and Pali (in Burmese script) text line recognition. Trained from scratch on 897,883 rendered line images with augmentation for real-world robustness.

Model Description

• Architecture: CRNN (4 conv layers + 3× bidirectional LSTM(200) + linear output)
• Parameters: 4.0M
• Alphabet: 118 characters (Burmese consonants, vowels, tones, medials, digits, Pali combinations, punctuation)
• Input: grayscale/bi-level line image, height normalized to 120px
• Output: CTC-decoded Burmese/Pali text

Performance

Evaluated on 10 held-out page images (6 modern Burmese, 4 Pali/nissaya):

Category	CER
Overall	0.77%
Modern Burmese	0.92%
Pali/nissaya	0.60%

Training Details

• Training data: 718,515 lines (80% clean, 20% degraded with skew, speckle, blur, strokes, page-curve, perspective)
• Dataset: pndaza/burmese-kraken-1m
• Optimizer: AdamW, learning rate 0.001 → 0.0001 (fine-tuning phase)
• Precision: 16-mixed (FP16)
• Hardware: NVIDIA RTX 5090 (vast.ai)
• Val accuracy: 0.9785

Usage

Command-line (kraken)

# Download the model
hf download pndaza/burmese-kraken-ocr burmese_kraken_ocr.safetensors

# Run OCR on a page image
kraken -i page.png output.txt binarize segment ocr -m burmese_kraken_ocr.safetensors

# For dense/degraded images, use the neural baseline segmenter
kraken -i page.png output.txt binarize segment -bl ocr -m burmese_kraken_ocr.safetensors

Python (kraken API)

from kraken import blla
from kraken.lib import vgsl
from kraken.lib.util import open_image

# Load model
model = vgsl.TorchVGSLModel.load_model("burmese_kraken_ocr.safetensors")

# Segment + recognize
im = open_image("page.png")
seg = blla.segment(im)
from kraken import rpred
recognizer = rpred.rpred(model, im, seg)
result = next(recognizer)
print(result)

Notes

• This is a recognition model only. Segmentation uses Kraken's default blla.mlmodel baseline segmenter, which works well for most documents but may produce errors on dense Burmese text with deep descenders.
• Line images are rendered, not scanned. The training data includes degraded augmentations to bridge the domain gap, but performance on real scans may vary.
• For the general-purpose image dataset (not Kraken format), see pndaza/burmese-ocr-1m.