Powered by PyTorch • Optimized for Banking & Security Systems • Research-Grade Quality
- ✨ Features
- 🚀 What's New in v2.0
- 🏗️ Architecture
- ⚡ Quick Start
- 📊 Performance Metrics
- 🧪 Testing
- 🔬 Technical Deep Dive
- 📚 State-of-the-Art References (2024-2025)
- 🛠️ Tech Stack
- 🗂️ Project Structure
- 🎓 Research & Citations
- 🤝 Contributing
- 📜 License
| 🎯 Feature | 📝 Description |
|---|---|
| 🧠 Deep Learning | Siamese Convolutional Neural Network with Contrastive Loss |
| ⚡ High Accuracy | Optimized for banking-grade precision-recall balance |
| 🔄 Transfer Learning | Pre-trained models ready for fine-tuning |
| 📊 ROC Analysis | Comprehensive evaluation with ROC curves |
| 🎨 Production-Ready | Modern package structure, tested, and documented |
| 🧪 Fully Tested | 44 comprehensive tests with 100% pass rate |
| 📦 Installable Package | pip install -e . for easy development |
| 🛠️ Modern Tooling | Black, Ruff, mypy, pytest configured |
| 📝 Type Hints | Complete type coverage for safety |
| 🚀 GPU Accelerated | Full CUDA support for faster training |
Version 2.0 is a complete rewrite with modern Python packaging, comprehensive testing, and production-grade code quality.
- ✅ Modern Package Structure: src/ layout with proper imports
- ✅ Comprehensive Testing: 44 tests covering all components
- ✅ Type Safety: Full type hints with mypy validation
- ✅ Development Tools: Black, Ruff, pytest, pre-commit hooks
- ✅ Production Scripts: CLI-ready with argparse
- ✅ Complete Documentation: CHANGELOG, CONTRIBUTING, LESSONS-LEARNED
# OLD (v1.x) - Direct file imports
from Model import SiameseConvNet
# NEW (v2.0) - Package imports
from signature_verification import SiameseConvNetSee CHANGELOG.md for full migration guide.
graph LR
A[🔷 Siamese CNN] --> B[🔶 Vision Transformers]
B --> C[🔷 Hybrid CNN-ViT]
C --> D[🌟 SOTA 2024-2025]
style A fill:#e1f5ff
style B fill:#fff3e0
style C fill:#f3e5f5
style D fill:#e8f5e9
- ✅ Siamese Convolutional Network - Proven architecture with contrastive learning
- ✅ PyTorch Framework - Modern, flexible, and production-ready
- ✅ Banking-Optimized - High recall for fraud detection
| Technology | Status | Impact |
|---|---|---|
| 🤖 Vision Transformers (ViT) | 📋 Planned | Global feature extraction |
| 🎯 Swin Transformers | 📋 Planned | Hierarchical attention mechanisms |
| ⚡ Hybrid CNN-ViT | 📋 Planned | Best of both worlds |
| 🔄 Few-Shot Learning | 📋 Planned | Learn from limited samples |
| 🎨 Spatial Transformer Networks | 📋 Planned | Automatic alignment |
┌─────────────────────────────────────────────────────────────┐
│ SIAMESE NETWORK │
├─────────────────────────────────────────────────────────────┤
│ │
│ Input Signature 1 Input Signature 2 │
│ │ │ │
│ ▼ ▼ │
│ ┌─────────┐ ┌─────────┐ │
│ │ Conv1 │ │ Conv1 │ │
│ │ 11x11 │ │ 11x11 │ Shared │
│ └────┬────┘ └────┬────┘ Weights │
│ │ │ │
│ ┌────▼────┐ ┌─────▼───┐ │
│ │ Pool + │ │ Pool + │ │
│ │ LRN │ │ LRN │ │
│ └────┬────┘ └────┬────┘ │
│ │ │ │
│ ┌────▼────┐ ┌─────▼───┐ │
│ │ Conv2 │ │ Conv2 │ │
│ │ 5x5 │ │ 5x5 │ │
│ └────┬────┘ └────┬────┘ │
│ │ │ │
│ ┌────▼────┐ ┌─────▼───┐ │
│ │ Conv3-4 │ │ Conv3-4 │ │
│ │ Dropout │ │ Dropout │ │
│ └────┬────┘ └────┬────┘ │
│ │ │ │
│ ┌────▼────┐ ┌─────▼───┐ │
│ │ FC │ │ FC │ │
│ │ 128-dim │ │ 128-dim │ │
│ └────┬────┘ └────┬────┘ │
│ │ │ │
│ └──────────┬───────────────┘ │
│ ▼ │
│ ┌──────────────┐ │
│ │ Euclidean │ │
│ │ Distance │ │
│ └──────┬───────┘ │
│ ▼ │
│ ┌──────────────┐ │
│ │ Contrastive │ │
│ │ Loss │ │
│ └──────────────┘ │
│ │
└─────────────────────────────────────────────────────────────┘
🔹 Convolutional Layers: Extract local features
🔹 Local Response Normalization: Enhance contrast
🔹 MaxPooling: Spatial dimensionality reduction
🔹 Dropout (0.3-0.5): Prevent overfitting
🔹 Fully Connected: 128-dimensional embeddings
🔹 Contrastive Loss: Metric learning optimization# Clone the repository
git clone https://github.com/umitkacar/Offline_Signature_Verification.git
cd Offline_Signature_Verification
# Create virtual environment (recommended)
python -m venv venv
source venv/bin/activate # On Windows: venv\Scripts\activate
# Install package in editable mode (recommended for development)
pip install -e .
# Or install from requirements.txt
pip install -r requirements.txt
# Install development dependencies (optional)
pip install -e ".[dev]"The package is now properly structured and installable:
# Install as editable package
pip install -e .
# Now import anywhere
python
>>> from signature_verification import SiameseConvNet, TrainDataset
>>> model = SiameseConvNet()
>>> print(model)# Step 1: Prepare data (first time only)
python scripts/prepare_data.py
# Step 2: Train the model
python scripts/train.py --epochs 5 --batch-size 8 --lr 0.001
# Advanced training options
python scripts/train.py \
--epochs 10 \
--batch-size 16 \
--lr 0.0005 \
--device cuda \
--model-dir ./MyModels# Evaluate trained model
python scripts/evaluate.py --model ./Models/checkpoint_epoch_4
# Custom evaluation
python scripts/evaluate.py \
--model ./Models/checkpoint_epoch_9 \
--data ./Data/test_index.pkl \
--output ./my_results.png \
--batch-size 16from signature_verification import (
SiameseConvNet,
SignatureTestDataset,
distance_metric
)
import torch
# Load model
model = SiameseConvNet()
model.load_state_dict(torch.load('./Models/checkpoint_epoch_4'))
model.eval()
# Load test data
dataset = SignatureTestDataset(data_path='./Data/test_index.pkl')
# Compare signatures
with torch.no_grad():
img1, img2, label = dataset[0]
img1 = img1.unsqueeze(0) # Add batch dimension
img2 = img2.unsqueeze(0)
features1, features2 = model(img1, img2)
distance = distance_metric(features1, features2)
is_genuine = distance < threshold # threshold = 1.5 (example)
print(f"Distance: {distance.item():.4f}")
print(f"Prediction: {'Genuine' if is_genuine else 'Forged'}")# Run quick functionality test (no data needed)
python scripts/quick_test.py| Metric | Value | Priority |
|---|---|---|
| 🎯 Recall (Sensitivity) | High | 🔴 Critical for fraud detection |
| 📊 Precision | Balanced | 🟡 Prevent customer inconvenience |
| 📈 F1-Score | Optimized | 🟢 Overall performance |
| ⚡ Inference Speed | < 50ms | 🔵 Real-time capability |
Banking Priority: Detecting forged signatures is critical! High recall prevents fraud while maintaining reasonable precision to avoid excessive customer re-verification.
┌─────────────────────────────────────────┐
│ Precision vs Recall Trade-off │
├─────────────────────────────────────────┤
│ │
│ High Recall → Catch more frauds │
│ (Priority) → Some false positives │
│ │
│ Balanced → Optimal UX │
│ Precision → Minimize re-signing │
│ │
└─────────────────────────────────────────┘
The Contrastive Loss function optimizes the network to:
- Minimize distance for genuine signature pairs
- Maximize distance for forged signature pairs
L = (1 - Y) × D² + Y × max(margin - D, 0)²
where:
Y = 1 (different persons) or 0 (same person)
D = Euclidean distance between embeddings
margin = 2.0Architecture:
Input: 220x155 grayscale images
Conv1: 48 filters, 11×11 kernel
Conv2: 128 filters, 5×5 kernel
Conv3: 256 filters, 3×3 kernel
Conv4: 96 filters, 3×3 kernel
FC1: 1024 neurons
FC2: 128-dimensional embeddings
Regularization:
- Dropout: 0.3 (after conv2 & conv4)
- Dropout: 0.5 (after fc1)
- Local Response Normalization
Optimization:
- Optimizer: Adam
- Learning Rate: 0.001
- Batch Size: 8| 📅 Year | 🔬 Research | 🏆 Highlights | 🔗 Link |
|---|---|---|---|
| 2025 | HTCSigNet | Hybrid Transformer-CNN, SOTA accuracy | 📄 |
| 2025 | PAST | Pairwise Attention Swin Transformer | 📄 |
| 2025 | Spatial Transformers | Automatic signature alignment | 📄 |
| 2025 | CNN-ViT Hybrid | 98.9% accuracy in biometrics | 📄 |
| 2024 | Vision Transformers | ViT market growth 33.2% CAGR | 📄 |
mindmap
root((2024-2025
Trends))
Vision Transformers
Swin Transformer
ViT Architecture
Hybrid CNN-ViT
Few-Shot Learning
Meta-Learning
N-way K-shot
Prototypical Networks
Self-Attention
Global Context
Multi-Head Attention
Cross-Attention
Advanced Architectures
Spatial Transformers
CycleGAN Denoising
YOLOv5 Detection
-
SigNet (2017) - Original Siamese CNN architecture
- Paper: Learning Features for Offline Handwritten Signature Verification
- Original: TensorFlow Implementation
- PyTorch: Community Port
-
Vision Transformers (2020) - Transformer architecture for CV
- Revolutionizing computer vision with attention mechanisms
- 280M to 2.7B USD market growth (2024-2032)
-
Contrastive Learning - Metric learning fundamentals
🔹 PyTorch 2.0+ - Dynamic computation graphs 🔹 CUDA Support - GPU acceleration 🔹 Mixed Precision - Faster training 🔹 DataLoader - Efficient batch processing 🔹 Model Checkpointing - Save/resume training
| 🏦 Industry | 💡 Application | 🎯 Benefit |
|---|---|---|
| Banking | Check verification | Prevent fraud |
| Legal | Document authentication | Ensure validity |
| Healthcare | Prescription verification | Patient safety |
| Government | ID verification | Security enhancement |
| Finance | Contract validation | Legal compliance |
Offline_Signature_Verification/
│
├── 📁 src/signature_verification/ # Main package (installable)
│ ├── 📄 __init__.py # Package exports
│ ├── 📄 model.py # Siamese CNN & Contrastive Loss
│ ├── 📄 dataset.py # PyTorch Dataset classes
│ └── 📄 utils.py # Utility functions
│
├── 📁 scripts/ # Production-ready scripts
│ ├── 📄 prepare_data.py # Data preprocessing
│ ├── 📄 train.py # Training with CLI args
│ ├── 📄 evaluate.py # Evaluation & ROC curves
│ ├── 📄 quick_test.py # Functionality verification
│ └── 📄 README.md # Scripts documentation
│
├── 📁 tests/ # Comprehensive test suite
│ ├── 📄 test_model.py # Model tests (25 tests)
│ ├── 📄 test_dataset.py # Dataset tests (8 tests)
│ ├── 📄 test_utils.py # Utils tests (11 tests)
│ └── 📄 conftest.py # Pytest configuration
│
├── 📁 Data/ # Training/test indices (gitignored)
├── 📁 Data_raw/ # Raw signature images (gitignored)
├── 📁 Models/ # Saved checkpoints (gitignored)
│
├── 📄 pyproject.toml # Modern build configuration
├── 📄 requirements.txt # Dependencies
├── 📄 .pre-commit-config.yaml # Pre-commit hooks
├── 📄 CHANGELOG.md # Version history
├── 📄 CONTRIBUTING.md # Developer guide
├── 📄 LESSONS-LEARNED.md # Refactoring insights
├── 📄 README.md # This file
├── 📄 LICENSE # MIT License
└── 📄 .gitignore # Git ignore rules
- src/ layout: Modern Python packaging standard
- Installable package:
pip install -e .for development - Type hints: Full type coverage with mypy
- Comprehensive tests: 44 tests with 100% pass rate
- Production scripts: CLI-ready with argparse
- Development tools: Black, Ruff, mypy, pytest configured
-
🔥 EndToEnd Signature System (2024)
- YOLOv5 Detection + CycleGAN Cleaning + Verification
- GitHub
- Tech: YOLOv5, CycleGAN, PyTorch/TensorFlow
-
⭐ sigver - Feature Extraction Package
- Writer-dependent classifiers
- GitHub
- Tech: PyTorch, Pre-trained models
-
🎯 Signature Recognition
- Digital image processing + Neural networks
- GitHub
- Tech: OpenCV, TensorFlow, 201+ stars
-
🤗 Hugging Face Signature Detection
- Production-ready model serving
- Hugging Face
- Tech: Triton Server, ONNX, TensorRT
Dataset Statistics:
Total Persons: 79
Signatures per Person: 12
Training Samples: 20,000 (10K positive + 10K negative)
Test Split: 5%
Image Size: 220 × 155 pixels
Format: Grayscale PNGData_raw/genuines/
└── NFI-{person:03d}{sign:02d}{person:03d}.png
Example: NFI-001-01-001.png
Person 1, Signature 1, Person 1
ROC Curves are used because:
- ✅ Summarize TPR vs FPR trade-offs
- ✅ Effective for balanced datasets
- ✅ Probability threshold visualization
PR Curves are preferred when:
- ✅ Imbalanced datasets (more genuine than forged)
- ✅ Focus on positive class performance
- ✅ Banking applications (high recall priority)
We welcome contributions! Here's how you can help:
- 🐛 Report bugs and issues
- 💡 Suggest new features or improvements
- 📝 Improve documentation
- 🔬 Share research papers and implementations
- 🚀 Submit pull requests
- Fork the repository
- Create your feature branch (
git checkout -b feature/AmazingFeature) - Commit your changes (
git commit -m 'Add some AmazingFeature') - Push to the branch (
git push origin feature/AmazingFeature) - Open a Pull Request
- 🤖 Integrate Vision Transformer (ViT) architecture
- 🎯 Implement Swin Transformer for hierarchical features
- ⚡ Add Hybrid CNN-ViT model
- 🔄 Few-shot learning capabilities
- 🎨 Spatial Transformer Networks for alignment
- 📱 Mobile deployment (ONNX/TensorRT)
- 🌐 Web demo with FastAPI
- 📊 Comprehensive benchmarking suite
- 🔒 Security hardening and adversarial robustness
- 📚 Extended documentation and tutorials
This project is licensed under the MIT License - see the LICENSE file for details.
MIT License - Free for commercial and private use
Special thanks to:
- 🎓 Research Community - For advancing signature verification
- 🔬 Paper Authors - SigNet, Vision Transformers, HTCSigNet teams
- 💻 Open Source Contributors - PyTorch, scikit-learn communities
- 📊 Dataset Providers - Kaggle signature datasets
- 🌟 GitHub Community - For feedback and contributions
If you find this project useful, please consider giving it a ⭐!
© 2024-2025 Offline Signature Verification Project
Empowering secure authentication with deep learning
We maintain a robust test suite with 44 tests covering all components:
# Run all tests
pytest
# Run with coverage report
pytest --cov=src --cov-report=html
# Run in parallel (faster)
pytest -n auto
# Run specific test file
pytest tests/test_model.py -v| Component | Tests | Coverage |
|---|---|---|
| Model | 25 tests | Architecture, forward/backward, loss functions |
| Dataset | 8 tests | Loading, preprocessing, data handling |
| Utils | 11 tests | Image processing, tensor conversion |
| Total | 44 tests | 100% passing ✅ |
No data required - perfect for CI/CD:
python scripts/quick_test.pyOutput:
============================================================
QUICK FUNCTIONALITY TEST
============================================================
✅ All imports successful!
✅ Model initialized
✅ Forward pass successful!
✅ Loss calculation successful!
✅ Distance metric successful!
✅ Gradient computation successful!
ALL TESTS PASSED! ✅
============================================================
# Clone and install
git clone https://github.com/umitkacar/Offline_Signature_Verification.git
cd Offline_Signature_Verification
# Install with dev dependencies
pip install -e ".[dev]"
# Install pre-commit hooks
pre-commit install# Format code
black src tests scripts
# Lint code
ruff check src tests scripts --fix
# Type checking
mypy src
# Run all quality checks
pre-commit run --all-filesAutomatically run on every commit:
- Trailing whitespace removal
- End-of-file fixer
- YAML/JSON/TOML validation
- Black formatting
- Ruff linting
- mypy type checking
- pytest tests
- README.md: Main documentation (you are here)
- CHANGELOG.md: Version history and migration guides
- CONTRIBUTING.md: Development guidelines
- LESSONS-LEARNED.md: Refactoring insights
- scripts/README.md: Script usage guide
All code is fully documented with Google-style docstrings:
def forward(self, x: Tensor, y: Tensor) -> Tuple[Tensor, Tensor]:
"""Forward pass through both branches of the Siamese network.
Args:
x: First signature tensor of shape (batch_size, 1, 220, 155)
y: Second signature tensor of shape (batch_size, 1, 220, 155)
Returns:
Tuple of feature embeddings (f_x, f_y), each of shape (batch_size, 128)
"""-
Install the package:
pip install -e . -
Update imports:
# Old from Model import SiameseConvNet from Dataset import TrainDataset # New from signature_verification import SiameseConvNet, TrainDataset
-
Update script calls:
# Old python train_model.py # New python scripts/train.py
-
Run tests:
pytest
See CHANGELOG.md for complete migration instructions.