sigverse 0.1.0

A professional signature verification and embedding library using a Siamese EfficientNet-B0 model.

🖋️ SigVerse: Enterprise-Grade Signature Intelligence

SigVerse is a high-performance, developer-centric Python SDK for offline signature verification and forensic embedding generation. Built on modern deep learning architectures, it provides the "First Line of Defense" for document-heavy industries like FinTech, LegalTech, and Insurance.

SigVerse doesn't just "match" images; it understands the latent spatial features of a signature to distinguish between genuine handwriting and sophisticated forgeries.

---

🚀 Key Features

• Siamese Deep Learning Architecture: Trained on the GPDS-150 and custom datasets to learn the underlying physics of handwriting.
• EfficientNet-B0 Backbone: Optimized for the perfect balance between high-precision feature extraction and low-latency inference.
• Forensic Vector Embeddings: Generate 256-dimensional "signature fingerprints" for storage, fast search, and clustering.
• Production-Ready with ONNX: Decoupled from heavy deep learning frameworks; runs at blazing speeds on CPU or GPU via ONNXRuntime.
• Zero-Config Integration: Pre-packaged model weights included. Install and verify in seconds.

---

🏗️ The Engineering Behind SigVerse

Why EfficientNet-B0?

In a production KYC pipeline, latency is the ultimate bottleneck. We selected EfficientNet-B0 because it utilizes Compound Scaling—balancing depth, width, and resolution. This allows SigVerse to capture fine-grained pen-stroke details without the computational overhead of heavy Vision Transformers (ViTs).

The Siamese Advantage

Unlike standard classification models, SigVerse uses a Siamese Network. During training, the model learns to minimize the distance between genuine signature pairs (Anchors and Positives) while maximizing the distance from forgeries (Negatives) using Triplet Margin Loss.

Result: A model that understands identity rather than just appearance.

---

🛠️ Installation & Setup

SigVerse is designed to be a standard Python package.

1. Install via Pip

# From PyPI (Coming Soon)
pip install sigverse

# Or directly from the repository for development
pip install git+https://github.com/parikhshyamal1993/SigVerse.git

2. Standard Dependencies

• Python 3.9+
• OpenCV
• NumPy
• ONNXRuntime

---

💻 Quick Start

Verification: Is it a Match?

Verify if two signatures belong to the same individual with a single call.

from sigverse import SignatureVerifier

# Initialize verifier
verifier = SignatureVerifier()

# Verify two images (supports file paths or OpenCV arrays)
result = verifier.verify("anchor_genuine.png", "target_sample.jpg")

print(f"Matched: {result['matched']}")
print(f"Similarity Score: {result['match_percentage']}")
print(f"L2 Distance: {result['distance']}")

Embeddings: Generate Signature Fingerprints

Convert signatures into numerical vectors for database storage or similarity search.

from sigverse import SignatureEmbedder

embedder = SignatureEmbedder()
vector = embedder.embed("signature_sample.png")

# 'vector' is a 256-d normalized numpy array
print(vector.shape) # (256,)

---

📈 Performance & Calibration

Our latest test suite (running on 1487+ unique users) shows consistent accuracy across various handwriting styles.

Scenario	Expected	Accuracy
Genuine vs. Genuine	Match	~92%
Genuine vs. Forged	No Match	~88%
Different User	No Match	~99%

Note: You can calibrate the threshold (default=1.0) in the verify() method to tune for False Acceptance Rate (FAR) vs False Rejection Rate (FRR) based on your use case.

---

🔮 Future Roadmap (The KYC Hub)

SigVerse is Component #1 of the KYC Intelligence Hub. Upcoming enhancements include:

1. PAD (Presentation Attack Detection): Detecting screen spoofing (taking photos of signatures on a monitor).
2. Moiré Pattern Detection: Identifying digital interference in fraudulent captures.
3. Signature Extraction: Auto-detecting and cropping signatures from complex multi-page documents (ID cards, bank statements).

---

📄 License

This project is licensed under the MIT License.

---

Built by Shyamal Parikh Transforming Document AI for the Real World.