---
language: bn
license: mit
tags: [ocr, bengali, crnn, easyocr, ctc]
datasets: [mnsm92/bengali-ocr-dataset-1m]
metrics: [cer, wer]
---

# Bengali CRNN OCR — Custom EasyOCR Recognition Model

**DocReader BD — CSC4233 NLP Final Project, AIUB**

## Results

| Model | CER ↓ | WER ↓ | Char Accuracy |
|---|---|---|---|
| Tesseract (baseline) | ~0.45 | ~0.60 | ~55% |
| EasyOCR default | ~0.25 | ~0.40 | ~75% |
| **BengaliCRNN (ours)** | **0.0348** | **0.1020** | **96.5%** |

## Architecture
ResNet34 (grayscale) + 2× BiLSTM (hidden=256) + CTC loss
Vocab: 152 Bengali + English chars | Input: 64×200px

## Files
- `bengali_crnn.pth` — EasyOCR-ready weights (module. prefix)
- `phase1_best.pth` — clean weights for further training
- `bengali_crnn.py` — EasyOCR network definition
- `bengali_crnn.yaml` — EasyOCR config
- `vocab.json` — character vocabulary
- `config.json` — model config

## Usage
```python
import easyocr
reader = easyocr.Reader(
    lang_list=["bn"],
    recog_network="bengali_crnn",
    model_storage_directory="./bengali_ocr_model",
    user_network_directory="./bengali_ocr_model",
    gpu=True
)
results = reader.readtext("bengali_doc.jpg")
```