Sarjinkhan2003/bengali-ocr-recognition

Bengali OCR — Lightweight Recognition Model

Project: DocReader BD — CSC4233 NLP, AIUB Architecture: LightCNN + BiLSTM + CTC (~4.5M params) Training data: 1,000,000 Bengali word images

Results

Model	CER ↓	WER ↓	Params
Tesseract (bn)	~45%	~60%	—
EasyOCR default (bn)	~25%	~40%	~6M
TrOCR-base-printed (fine-tuned)	~8%	~15%	330M
BengaliCRNN (ours)	0.62%	2.95%	~4.5M

Quick start

# pip install huggingface_hub torch torchvision Pillow
from huggingface_hub import hf_hub_download
import importlib.util, json, torch
from torchvision import transforms
from PIL import Image

# 1. Download files from hub
repo = "Sarjinkhan2003/bengali-ocr-recognition"
net_path   = hf_hub_download(repo, "bengali_crnn.py")
ckpt_path  = hf_hub_download(repo, "bengali_crnn.pth")
vocab_path = hf_hub_download(repo, "vocab.json")

# 2. Load model
spec = importlib.util.spec_from_file_location("bengali_crnn", net_path)
mod  = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)

vocab    = json.load(open(vocab_path, encoding="utf-8"))
idx2char = {int(k): v for k, v in vocab["idx2char"].items()}
model    = mod.Model(1, 256, 256, vocab["num_classes"])
ckpt     = torch.load(ckpt_path, map_location="cpu")
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

# 3. Run inference
tf = transforms.Compose([
    transforms.Grayscale(1),
    transforms.Resize((64, 256)),
    transforms.ToTensor(),
    transforms.Normalize([0.5],[0.5])
])
img    = Image.open("word.jpg").convert("RGB")
tensor = tf(img).unsqueeze(0)
with torch.no_grad():
    out = model(tensor)
_, preds = out.permute(1,0,2).max(2)
chars, prev = [], None
for p in preds[0].tolist():
    if p != 0 and p != prev:
        chars.append(idx2char.get(p, ""))
    prev = p
print("".join(chars))

EasyOCR integration

import easyocr
reader = easyocr.Reader(
    lang_list=["bn"],
    recog_network="bengali_crnn",
    model_storage_directory="./model_dir",
    user_network_directory="./model_dir",
    gpu=True
)
results = reader.readtext("bengali_doc.jpg")
for bbox, text, confidence in results:
    print(f"{confidence:.2f} | {text}")

Files

File	Description
bengali_crnn.pth	Model weights
bengali_crnn.py	Network architecture (EasyOCR compatible)
vocab.json	Bengali+English vocabulary (148 chars)
inference.py	Standalone inference helper
training_curves.png	Loss/CER/WER curves

Vocabulary

Bengali vowels, consonants, diacritics (incl. matra, hasanta, anusvar) + Bengali numerals + English letters/digits + punctuation Total: 148 characters