handwriting-recognition-iam

Handwriting Recognition (CNN-BiLSTM-CTC)

A complete handwriting recognition system using a CNN-BiLSTM-CTC architecture trained on the IAM handwriting database. Achieves 12.95% CER and 42.47% WER on the IAM test set.

Hugging Face: IsmatS/handwriting_recognition

Quick Start

pip install torch torchvision pillow numpy huggingface_hub

import torch
import torch.nn as nn
import numpy as np
from PIL import Image
from huggingface_hub import hf_hub_download

# Download model checkpoint
ckpt_path = hf_hub_download(repo_id="IsmatS/handwriting-recognition-iam", filename="best_model.pth")
checkpoint = torch.load(ckpt_path, map_location="cpu")

# Character mapper from checkpoint
char_mapper = checkpoint['char_mapper']

def preprocess_image(image_path, target_height=64, target_width=256):
    img = Image.open(image_path).convert('L')  # grayscale
    img = img.resize((target_width, target_height), Image.LANCZOS)
    img = np.array(img, dtype=np.float32) / 255.0
    img = (img - 0.5) / 0.5  # normalize to [-1, 1]
    return torch.FloatTensor(img).unsqueeze(0).unsqueeze(0)  # (1, 1, H, W)

def ctc_decode(predictions, char_mapper):
    """Greedy CTC decoding."""
    pred_indices = predictions.argmax(dim=2).squeeze(1).tolist()
    decoded = []
    prev = None
    for idx in pred_indices:
        if idx != prev and idx != 0:  # 0 = blank token
            decoded.append(char_mapper.idx_to_char[idx])
        prev = idx
    return ''.join(decoded)

# Note: CRNN class must match the training definition in train_colab.ipynb
# See train_colab.ipynb for the full model class
# model = CRNN(num_chars=len(char_mapper.chars))
# model.load_state_dict(checkpoint['model_state_dict'])
# model.eval()
#
# img_tensor = preprocess_image("handwriting_sample.png")
# with torch.no_grad():
#     output = model(img_tensor)  # (T, 1, num_chars)
# text = ctc_decode(output, char_mapper)
# print("Recognized:", text)

Note: The trained model weights (best_model.pth) are generated during training. Run train_colab.ipynb on Google Colab to produce the checkpoint, then use the code above for inference.

📁 Files

1. analysis.ipynb - Dataset Analysis

• Exploratory Data Analysis (EDA)
• 5 detailed charts saved to charts/ folder
• Run locally or on Colab (no GPU needed)

2. train_colab.ipynb - Model Training (GPU)

• ⚡ Google Colab GPU compatible
• Full training pipeline
• CNN-BiLSTM-CTC model (~9.1M parameters)
• Automatic model saving
• Download trained model for deployment

🚀 Quick Start

Option 1: Analyze Dataset (Local/Colab)

jupyter notebook analysis.ipynb

• No GPU needed
• Generates 5 EDA charts
• Fast (~2 minutes)

Option 2: Train Model (Google Colab GPU)

1. Upload train_colab.ipynb to Google Colab
2. Change runtime to GPU:
   • Runtime → Change runtime type → GPU (T4 recommended)
3. Run all cells
4. Download trained model (last cell)

Training Time: ~1-2 hours for 20 epochs on T4 GPU

📊 Charts Generated

From analysis.ipynb:

1. charts/01_sample_images.png - 10 sample handwritten texts
2. charts/02_text_length_distribution.png - Text statistics
3. charts/03_image_dimensions.png - Image analysis
4. charts/04_character_frequency.png - Character distribution
5. charts/05_summary_statistics.png - Summary table

🎯 Model Details

Architecture:

• CNN: 7 convolutional blocks (feature extraction)
• BiLSTM: 2 layers, 256 hidden units (sequence modeling)
• CTC Loss: Alignment-free training

Dataset: Teklia/IAM-line (Hugging Face)

• Train: 6,482 samples
• Validation: 976 samples
• Test: 2,915 samples

Metrics:

• CER (Character Error Rate)
• WER (Word Error Rate)

💾 Model Files

After training in Colab:

• best_model.pth - Trained model weights
• training_history.png - Loss/CER/WER plots
• predictions.png - Sample predictions

📦 Requirements

torch>=2.0.0
datasets>=2.14.0
pillow>=9.5.0
numpy>=1.24.0
matplotlib>=3.7.0
seaborn>=0.13.0
jupyter>=1.0.0
jiwer>=3.0.0

🔧 Usage

Load Trained Model

import torch

# Load checkpoint
checkpoint = torch.load('best_model.pth')
char_mapper = checkpoint['char_mapper']

# Create model
from train_colab import CRNN  # Copy model class
model = CRNN(num_chars=len(char_mapper.chars))
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Predict
# ... (preprocessing + inference)

📝 Notes

• GPU strongly recommended for training (use Colab T4)
• Training on CPU will be extremely slow (~20x slower)
• Colab free tier: 12-hour limit, sufficient for 20 epochs
• Model checkpoint includes character mapper for deployment

🎓 Training Tips

1. Start with fewer epochs (5-10) to test
2. Monitor CER/WER - stop if not improving
3. Increase epochs if still improving (up to 50)
4. Save checkpoint before Colab disconnects
5. Download model immediately after training

📄 License

Dataset: IAM Database (research use)