ne-ocr

<p align="center"> <img src="https://huggingface.co/MWirelabs/ne-ocr/resolve/main/assets/mwire.png" width="180" alt="MWire Labs Logo"> </p>

NE-OCR

High-Accuracy OCR for Northeast Indian Scripts

Purpose-built OCR for Northeast India with 94.99% average character accuracy across 12 language–script pairs.
Outperforms EasyOCR, Tesseract 5, and TrOCR-large on 9 of 12 language–script pairs.
Fast inference and strong performance where general OCR systems fail.

Developed by MWire Labs, Shillong, Meghalaya.

<p align="center"> <img src="https://huggingface.co/MWirelabs/ne-ocr/resolve/main/assets/neocrarchitecture.jpg" width="850" alt="NE-OCR Architecture Diagram"> </p>

NE-OCR is built on a ViTSTR-Base encoder with CTC decoding. The model processes 32×128 RGB word/line crops across Latin, Bengali, Devanagari, and Meitei Mayek scripts, outputting text from a 1,056-character multilingual vocabulary.

Model Details

• Architecture: DocTR ViTSTR-Base (86M parameters)
• Vocab size: 1056 characters (Latin, Bengali, Devanagari, Meitei Mayek)
• Input: 32×128 RGB image crops (word/line level, ≤32 chars)
• Training data: ~988k deduplicated samples across 12 languages
• Trained by: MWire Labs

Inference Speed

Measured on NVIDIA A40 (batch size = 1):

<p align="center"> <img src="https://huggingface.co/MWirelabs/ne-ocr/resolve/main/assets/inferenceneocr.png" width="700" alt="NE-OCR Latency Comparison"> </p>

• NE-OCR: 17.2 ms/image
• EasyOCR: 37.2 ms
• TrOCR-large: 92.1 ms
• Tesseract 5: 166.1 ms
• Chandra (VLM): 313 ms

NE-OCR is:

• 2× faster than EasyOCR
• 9× faster than Tesseract
• 18× faster than VLM-based OCR systems

Benchmark Comparison — Character Accuracy (ChA%)

Evaluated on a fixed 26,000-sample benchmark (2,000 per language–script pair).
Higher is better.

Benchmark Test Set

A public benchmark test set is available in the benchmark/ folder of this repository for reproducing evaluation results and comparing against other OCR models.

• Combined: benchmark/ne_ocr_benchmark.parquet — 26,000 samples across all 12 languages
• Per-language: benchmark/{lang}_test.parquet — 2,000 samples each
• Format: Parquet with columns: image_path, text, lang
• Filter: All samples ≤32 characters (word/line-level crops)

Results reported in this model card are computed on this exact test set.

Usage

import torch, json
import numpy as np
from PIL import Image
from huggingface_hub import hf_hub_download
from doctr.models import vitstr_base

# Download files
model_path = hf_hub_download(repo_id='MWirelabs/ne-ocr', filename='ne_ocr_best.pt')
vocab_path = hf_hub_download(repo_id='MWirelabs/ne-ocr', filename='ne_ocr_vocab.json')

# Load vocab
with open(vocab_path, encoding='utf-8') as f:
    vocab_data = json.load(f)
vocab_str = ''.join(vocab_data['vocab'][1:])

# Load model
model = vitstr_base(pretrained=False, vocab=vocab_str)
model.load_state_dict(torch.load(model_path, map_location='cpu'))
model.eval()

# Inference (word/line crop, max 32 chars)
img = Image.open('your_crop.jpg').convert('RGB').resize((128, 32))
img_tensor = torch.tensor(np.array(img, dtype=np.float32)/255.0).permute(2,0,1).unsqueeze(0)
out = model(img_tensor)
print(out['preds'][0][0])

Notes

• Model is designed for word/line-level crops (≤32 characters), not full pages
• For full page OCR, use a text detection model first (e.g. DBNet) to extract crops
• Bodo accuracy is lower due to limited training data; planned improvement in V2

License

CC-BY-4.0 — MWire Labs