🖋️ Signature Identification System (Siamese ResNet-based)

An intelligent signature verification and identification system built using a Siamese ResNet architecture.
It verifies whether a given signature belongs to a registered user by comparing deep feature embeddings of signature images.

This system is optimized for limited user data (as few as 5 samples per user), using data augmentation and contrastive learning for robust verification.

---

🚀 Features

✅ ResNet-18 based Siamese Network – extracts deep signature embeddings.
✅ Contrastive Loss Training – learns to distinguish between genuine and forged pairs.
✅ Automatic Embedding Enrollment – generates and saves student signature representations.
✅ Real-time Verification – compare uploaded signatures with enrolled ones.
✅ Augmentation Pipeline – increases dataset size for limited samples.
✅ Support for CEDAR + Student Dataset – seamlessly integrates both datasets.

---

🧩 Project Structure

Signature-Identification/
│
├── dataset.py                 # Loads and augments signature data
├── model.py                   # Defines Siamese ResNet model
├── train.py                   # Trains the model using CEDAR + student data
├── predict.py                 # Verifies a new signature against enrolled users
├── enroll.py                  # Scans test images and saves embeddings
│
├── signatures/                # Public signature dataset (already large)
│── students/                  # Contains student signatures (only 5 per user)
│       ├── student1/
│       ├── student2/
│       └── ...
│
├── models/
│   ├── student_embeddings.pkl # Saved student embeddings after enrollment
    └── siamese_resnet.pth   # Trained model weights

---

⚙️ Setup Instructions

1️⃣ Clone the repository

git clone https://github.com/yourusername/Signature-Identification.git
cd Signature-Identification

2️⃣ Install dependencies

pip install -r requirements.txt

---

🏋️‍♂️ Training the Model

Train the Siamese ResNet using both CEDAR and student signatures:

python train.py

This script:

• Loads and augments the datasets
• Creates genuine/forged signature pairs
• Trains the Siamese ResNet model with contrastive loss
• Saves model weights to models/siamese_resnet.pth

---

🧬 Enrollment (Generate Student Embeddings)

Once training is done, generate embeddings for all student signatures:

python enroll.py

This step:

• Loads the trained model
• Computes feature embeddings for each student's signatures
• Saves them in mdoels/student_embeddings.pkl

---

🔍 Signature Verification

Verify a given signature against all enrolled users, by configuring query_path:

python predict.py

The system compares the query signature with all stored embeddings and prints the most probable match, along with the similarity score.

Example Output:

Predicted Student: student5
Similarity Score: 0.912
Result: Genuine

---

🧪 Model Architecture Overview

The Siamese ResNet model uses two identical branches of ResNet-18 sharing weights.

        Input 1                  Input 2
           │                       │
           ▼                       ▼
      ResNet Encoder          ResNet Encoder
           │                        │
           └── Feature Embeddings ──┘
                     │
             Euclidean Distance
                     │
                Contrastive Loss

This allows the network to learn meaningful signature representations and distinguish genuine vs forged pairs effectively.

---

📈 Performance

Dataset	Accuracy	Loss	Remarks
CEDAR	~97%	Low	High-quality dataset
Students (5→Augmented)	~93%	Moderate	Improved with augmentation

---

🛠️ Technologies Used

• Python 3.11
• PyTorch – deep learning framework
• OpenCV / PIL – image handling
• Albumentations – data augmentation
• NumPy / Pandas – preprocessing utilities

---

🧾 Future Improvements

• ✅ Web Interface for Upload & Verification
• ✅ Add One-shot Learning for new users
• 🔲 Integrate Live Signature Capture

---