Forgery Detection - Signature Verification System

This project is a Signature Forgery Detection System built using Django and TensorFlow/Keras, which leverages transfer learning with pre-trained CNN models including ResNet50, VGG16, VGG19, and a custom CNN model. The aim is to classify whether a given signature is genuine or forged.

---

📌 Project Objectives

• Detect forged signatures using deep learning
• Use transfer learning with pre-trained models for efficient training
• Train multiple models for comparison and accuracy benchmarking
• Serve predictions through a Django web interface

---

🧠 Model Architectures

1. 🔧 Custom CNN Model

A standard CNN with multiple convolution and pooling layers followed by dense layers.

• Input size: 150x150
• Layers: Conv2D → MaxPooling → Dense
• Optimizer: Adam (lr=0.0001)
• Loss: Binary Crossentropy
• Metrics: Accuracy
• Epochs: 100 (with early stopping)

---

2. 🧠 ResNet50 (Transfer Learning)

• Pre-trained on ImageNet (include_top=False)
• Input size: 64x64
• Pooling: avg
• Added Layers: Flatten → Dense(512 → 435 → 365 → 1)
• Trainable Layers: All frozen
• Optimizer: Adam
• Epochs: 100 (early stopping with patience 30)

---

3. 🧠 VGG16 (Transfer Learning)

• Pre-trained on ImageNet (include_top=False)
• Input size: 64x64
• Pooling: avg
• Added Layers: Flatten → Dense(512 → 450 → 260 → 1)
• Trainable Layers: All frozen
• Optimizer: Adam
• Epochs: 100 (early stopping with patience 30)

---

4. 🧠 VGG19 (Transfer Learning)

• Pre-trained on ImageNet (include_top=False)
• Input size: 64x64
• Pooling: avg
• Added Layers: Flatten → Dense(512 → 455 → 250 → 1)
• Trainable Layers: All frozen
• Optimizer: Adam
• Epochs: 100 (early stopping with patience 30)

---

🧰 Technologies & Libraries Used

This project combines deep learning with web development to deliver a full-stack signature forgery detection system.

🔧 Backend

• Django — Web framework to handle routing, views, form submissions, file storage, and templating.
• Python — Core programming language for backend logic and model execution.

📄 Templating & Styling

• Django Template Language (DTL) — For dynamic HTML rendering with logic blocks, URL routing, and static file linking.
• Bootstrap 5 — For responsive layout, modern UI components, grid system, and modals.
• Bootstrap Icons — Used to visually enhance success/failure results.
• Custom CSS — For additional design and theming.

🤖 Machine Learning / Deep Learning

• TensorFlow / Keras — To train, load, and infer from deep learning models:
  • Custom Sequential CNN Model
  • ResNet50 (Transfer Learning)
  • VGG16 (Transfer Learning)
  • VGG19 (Transfer Learning)
• NumPy / OpenCV (optional) — For image preprocessing.

🌐 Web Functionality

• Signature image upload via Django form.
• Real-time model prediction for 4 different CNN architectures.
• Display of:
  • Confidence scores per model
  • Averaged result and final verification decision
  • Signature preview for the uploaded image
  • Responsive and mobile-friendly layout.

---

📂 Dataset

Dataset used: Signature Verification Dataset by robinreni

Downloaded using kagglehub:

import kagglehub

path = kagglehub.dataset_download("robinreni/signature-verification-dataset")
print("Path to dataset files:", path)

🔗 Download Pre-trained Models

To get started quickly, you can download the pre-trained .h5 models from the following Google Drive links:

Model Name	Architecture	Download Link
Custom CNN	Custom-built CNN	Download
ResNet50	Transfer Learning	Download
VGG16	Transfer Learning	Download
VGG19	Transfer Learning	Download

🚀 Getting Started

• Ensure you have installed Docker.

📦 Clone the Repository

git clone https://github.com/sanu0711/Forgery-Detection-Signature-Verification.git

🐳 Navigate to the project directory

cd Forgery-Detection-Signature-Verification

🛠️ Build and Run the Application

This command will build the Docker image and start the Django server inside a container.

• Place the downloaded .h5 files inside the trained_models/ directory before running the app.

docker-compose up --build

🌐 Access the Web App

http://localhost:8000/

Here, you can upload signature images and get real-time forgery predictions from 4 different deep learning models

🎥 Demo

---