kannada-parseq-v2

kannada-parseq-v2

Word-level Kannada scene/document text recognizer (PARSeq architecture), trained from scratch on 648k word images derived from public Karnataka government documents plus synthetic scan-degraded and numeral data.

Results

Training data

• 529k word crops from born-digital Karnataka government PDFs (gazettes, circulars, forms), labels recovered by converting legacy Nudi font encodings to Unicode — no manual annotation
• 120k synthetic samples: scan degradations (blur, noise, low-res, binarization) + numerals (dates, Aadhaar-format numbers, amounts, Kannada digits ೦–೯)
• Charset: 160 symbols (Kannada block + ASCII + ₹ + common punctuation)

Intended use & limitations

Reading printed Kannada in documents (government, banking, general print). Not for handwriting. Scanned-document accuracy trails born-digital. Max label length 42 chars; input 32×128 (resize word crops). Pair with a text detector (e.g. CRAFT) for full pages.

Trained on public government documents; may reflect the formal/administrative register more strongly than colloquial text.

Usage

# pip install torch timm pytorch-lightning
# clone https://github.com/baudm/parseq for the strhub package
from strhub.models.utils import load_from_checkpoint
from strhub.data.module import SceneTextDataModule
import torch

model = load_from_checkpoint("kannada-parseq-v2.ckpt").eval()
transform = SceneTextDataModule.get_transform(model.hparams.img_size)
img = transform(word_image_pil).unsqueeze(0)
with torch.inference_mode():
    preds, _ = model.tokenizer.decode(model(img).softmax(-1))
print(preds[0])

Training

PARSeq base (~24M params), 32 epochs + SWA, batch 384, lr 7e-4, one-cycle, single NVIDIA GB10 (DGX Spark), ~10 hours. normalize_unicode must be disabled for Indic charsets (the default NFKD+ASCII step deletes Kannada).