Signature Verification Module

A comprehensive Python package for signature verification and classification using deep learning models. This module supports both ResNet Siamese networks for verification and Vision Transformers (ViT) with ArcFace for classification.

Features

• Multiple Model Architectures:
  • ResNet50 Siamese network with contrastive loss for signature verification
  • Vision Transformer (ViT) with ArcFace loss for signature classification
• S3 Integration: Seamless integration with AWS S3 for model weights and dataset storage
• Flexible Configuration: Environment variable-based configuration system
• Complete Pipeline: From data preprocessing to model training and inference
• Production Ready: Optimized for both research and production deployment

Installation

Requirements

• Python 3.8+
• PyTorch 1.9+
• torchvision
• timm (for ViT models)
• scikit-image
• scikit-learn
• pandas
• PIL/Pillow
• boto3 (for S3 support)
• pydantic
• scipy

Install Dependencies

pip install torch torchvision torchaudio
pip install timm scikit-image scikit-learn pandas pillow boto3 pydantic scipy

Quick Start

Basic Usage

from signature import SignatureConfig, load_model, classify_signatures

# Configure the module
config = SignatureConfig(
    model_type="vit",
    num_classes=100,
    model_s3_path="s3://my-bucket/models/signature_vit.pth"
)

# Load a trained model
model = load_model()

# Classify signatures
results = classify_signatures(
    model=model,
    reference_path="path/to/reference/signatures",
    detected_path="path/to/detected/signatures"
)

print(f"Accepted signatures: {results['accepted_signatures']}")

Training a Model

from signature import train_model, create_dataset_csv

# Create a dataset CSV from image directory
create_dataset_csv(
    image_directory="path/to/signature/images",
    output_csv="dataset.csv",
    dataset_type="single"  # or "pairs" for ResNet
)

# Train a model
model, metrics = train_model(
    model_type="vit",
    dataset_path="dataset.csv",
    save_path="s3://my-bucket/models/trained_model.pth",
    num_classes=100,
    epochs=20
)

print(f"Test accuracy: {metrics['test_accuracy']:.4f}")

Configuration

The module supports configuration through environment variables or programmatically:

Environment Variables

# Model configuration
export SIGNATURE_MODEL_TYPE=vit
export SIGNATURE_NUM_CLASSES=100
export SIGNATURE_MODEL_S3_PATH=s3://my-bucket/models/model.pth

# AWS configuration
export AWS_ACCESS_KEY_ID=your_access_key
export AWS_SECRET_ACCESS_KEY=your_secret_key
export AWS_DEFAULT_REGION=us-east-2

# Training configuration
export SIGNATURE_BATCH_SIZE=32
export SIGNATURE_LEARNING_RATE=5e-5
export SIGNATURE_EPOCHS=15

Programmatic Configuration

from signature import SignatureConfig, set_config

config = SignatureConfig(
    model_type="vit",
    num_classes=100,
    batch_size=32,
    learning_rate=5e-5,
    aws_access_key_id="your_access_key",
    aws_secret_access_key="your_secret_key"
)

set_config(config)

Model Architectures

ResNet50 Siamese

Best for signature verification (determining if two signatures are from the same person):

from signature import ResNet50Siamese, train_resnet

# Create model
model = ResNet50Siamese(embedding_dim=512)

# Train model
model, metrics = train_resnet(
    dataset_path="pairs_dataset.csv",
    epochs=25,
    margin=17.5
)

Vision Transformer (ViT)

Best for signature classification (identifying which person signed):

from signature import SignatureViT, train_vit

# Create model
model = SignatureViT(num_classes=100, embedding_dim=512)

# Train model
model, metrics = train_vit(
    num_classes=100,
    dataset_path="single_image_dataset.csv",
    epochs=15
)

Data Preparation

Directory Structure

Organize your signature images in the following structure:

signatures/
├── person1/
│   ├── signature1.png
│   ├── signature2.png
│   └── signature3.png
├── person2/
│   ├── signature1.png
│   └── signature2.png
└── person3/
    ├── signature1.png
    ├── signature2.png
    └── signature4.png

Creating Dataset CSVs

from signature import create_dataset_csv

# For ViT (single image classification)
create_dataset_csv(
    image_directory="signatures/",
    output_csv="vit_dataset.csv",
    dataset_type="single"
)

# For ResNet (image pairs)
create_dataset_csv(
    image_directory="signatures/",
    output_csv="resnet_dataset.csv",
    dataset_type="pairs",
    pairs_per_person=50
)

S3 Integration

The module seamlessly works with S3 for storing models, datasets, and reference signatures:

from signature import S3Manager

# Create S3 manager
s3_manager = S3Manager()

# Download model weights
s3_manager.download_file(
    "s3://my-bucket/models/model.pth", 
    "local_model.pth"
)

# Upload trained model
s3_manager.upload_file(
    "local_model.pth",
    "s3://my-bucket/models/new_model.pth"
)

Advanced Usage

Custom Feature Extraction

from signature import get_fused_features, extract_deep_features

# Extract fused features (deep + HOG)
features = get_fused_features(model, "signature.png")

# Extract only deep features
deep_features = extract_deep_features(model, "signature.png")

Custom Training Loop

from signature import (
    create_model, ContrastiveLoss, create_dataloader, 
    get_config
)
import torch

config = get_config()

# Create model and loss
model = create_model(model_type="resnet")
criterion = ContrastiveLoss(margin=17.5)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)

# Create data loaders
train_loader, val_loader, test_loader = create_dataloader(
    "dataset.csv", 
    dataset_type="pairs",
    batch_size=32
)

# Custom training loop
for epoch in range(10):
    for img1, img2, labels in train_loader:
        optimizer.zero_grad()
        emb1, emb2 = model(img1, img2)
        loss = criterion(emb1, emb2, labels)
        loss.backward()
        optimizer.step()

API Reference

Core Functions

• load_model(): Load a trained model with optional S3 weights
• classify_signatures(): Classify detected signatures against references
• train_model(): Train a model with automatic type detection
• create_dataset_csv(): Create dataset CSV from image directory

Model Classes

• ResNet50Siamese: Siamese ResNet50 for signature verification
• SignatureViT: Vision Transformer with ArcFace for classification
• ContrastiveLoss: Contrastive loss for Siamese networks
• ArcFace: Angular margin loss implementation

Data Classes

• SignatureDataset: Dataset for single signature images
• SignaturePairDataset: Dataset for signature image pairs
• create_dataloader(): Create train/val/test data loaders

Utility Classes

• SignatureConfig: Configuration management
• S3Manager: S3 operations manager

Performance Tips

1. Use GPU: Set SIGNATURE_DEVICE=cuda for faster training and inference
2. Batch Processing: Use larger batch sizes for better GPU utilization
3. Data Augmentation: Enable augmentation for better generalization
4. Caching: Models and datasets are automatically cached locally
5. S3 Optimization: Use appropriate S3 region for reduced latency

Troubleshooting

Common Issues

1. Import Errors: Ensure all dependencies are installed
2. S3 Access: Check AWS credentials and permissions
3. Memory Issues: Reduce batch size or use CPU if GPU memory is limited
4. Image Format: Ensure images are in supported formats (PNG, JPG, JPEG)

Debug Mode

Enable detailed logging:

from signature.utils import setup_logging
setup_logging("DEBUG")

Examples

See the examples/ directory for complete examples:

• train_vit_example.py: Training a ViT model
• train_resnet_example.py: Training a ResNet Siamese model
• inference_example.py: Running inference on new signatures
• s3_integration_example.py: Working with S3 storage

License

This module is part of the LiteWave ML Models repository.