Indra OCR Vietnamese (v1)

Fine-tuned version of datalab-to/indra-ocr optimized for high-accuracy Vietnamese document text recognition.

This checkpoint is a drop-in replacement for the official indra-ocr pipeline. It retains full architecture compatibility with Datalab's ONNX runtime while significantly boosting accuracy on Vietnamese administrative decrees, academic research publications, and structured forms.

Benchmark & Evaluation

Evaluated on held-out Vietnamese test data across administrative (Chinhphu), academic (VJOL), and real-world forms (Datalab snapshot):

Metric	Base Model (`datalab-to/indra-ocr`)	Fine-Tuned (`ANZ-Innovation/indra-v1`)	Improvement
Character Error Rate (CER NFD)	8.48%	1.49%	-82.44% relative
Exact Line Accuracy	39.44%	82.45%	+43.02% absolute

Training Stats & Efficiency

Hardware: 1x NVIDIA RTX A4000 16GB
Throughput: ~668.0 samples/second
Total Training Duration: ~~2419.8s (~~40.3 minutes, 16 epochs, best at epoch 14)
Dataset Partitioning:
- Train: 57,030 lines
- Validation: 3,887 lines
- Test (Held-out): 3,887 lines
- Total: 64,804 lines
- Sources:
  - Government decrees & administrative documents (Chinhphu): 3,445 lines
  - Academic journals & scientific publications (VJOL): 5,118 lines
  - Datalab DBNet detector-matched clean document lines: 21,241 lines
  - Photorealistic synthetic document lines (Liberation Serif/Sans, Lato, DejaVu): 35,000 lines

Quickstart & Usage

You can use this fine-tuned model directly with the official indra-ocr package or ONNX Runtime:

pip install indra-ocr huggingface_hub

1. In Python with `indra_ocr`:

from huggingface_hub import snapshot_download
import indra_ocr
import cv2

# Download model weights from Hugging Face
model_dir = snapshot_download(repo_id="ANZ-Innovation/indra-v1")

# Load OCR pipeline
converter = indra_ocr.load(path=model_dir)

# Read image (grayscale)
img = cv2.imread("document_vn.png", cv2.IMREAD_GRAYSCALE)

# Run full OCR (detection + recognition)
result = converter.ocr(img)

for line in result.lines:
    print(f"  {line.text}")

Technical Details

Recognizer Architecture: Multi-scale 1D depthwise separable convolutional backbone with 128-channel CTC classifier head (~1.24M parameters).
Format: High-performance ONNX model (rec.onnx) with frozen batch normalization and dynamic sequence length masking.
License: Modified OpenRAIL-M (commercial use permitted subject to Indra OCR terms).

Developed by ANZ Innovation.