Memory-Aware Automatic Differentiation for Optimization-Based Image Super-Resolution

📌 Overview

This project presents a memory-efficient image super-resolution framework that combines deep learning and optimization techniques to reconstruct high-resolution images from low-resolution inputs.

The proposed approach utilizes:

• FSRCNN (Fast Super-Resolution Convolutional Neural Network) for efficient image reconstruction.
• Automatic Differentiation (AD) for gradient-based image refinement.
• Conjugate Gradient (CG) optimization to improve convergence and enhance image quality.

The objective is to generate high-quality images while minimizing computational and memory requirements, making the solution suitable for resource-constrained environments.

---

🚀 Features

• Lightweight and memory-efficient super-resolution framework.
• Deep learning-based image reconstruction using FSRCNN.
• Gradient optimization using Automatic Differentiation.
• Image refinement using Conjugate Gradient optimization.
• Performance evaluation using PSNR and SSIM metrics.
• Suitable for medical imaging, surveillance, satellite imaging, and mobile photography applications.

---

🛠️ Tech Stack

• Python
• PyTorch
• NumPy
• OpenCV
• Matplotlib
• Scikit-Image
• Jupyter Notebook

---

📂 Project Structure

text Memory-Aware-Image-Super-Resolution/ │ ├── dataset/ ├── notebooks/ ├── models/ │ └── fsrcnn_model.py ├── optimization/ │ ├── automatic_differentiation.py │ └── conjugate_gradient.py ├── outputs/ ├── images/ ├── requirements.txt └── README.md

---

🔬 Methodology

Step 1: Data Preprocessing

• Load high-resolution images.
• Generate low-resolution images using bicubic downsampling.
• Resize and normalize images.

Step 2: Model Training

• Train the FSRCNN model on paired LR-HR images.

Step 3: Automatic Differentiation

• Compute gradients efficiently.
• Minimize reconstruction loss.

Step 4: Conjugate Gradient Optimization

• Refine the generated image.
• Improve convergence and preserve image details.

Final Pipeline

Low Resolution Image
        ↓
      FSRCNN
        ↓
Automatic Differentiation
        ↓
Conjugate Gradient
        ↓
High Resolution Output

---

📊 Evaluation Metrics

Peak Signal-to-Noise Ratio (PSNR)

Measures reconstruction quality.

Structural Similarity Index (SSIM)

Measures structural similarity between the original and reconstructed images.

---

📈 Results

The proposed framework achieves:

• Improved PSNR and SSIM scores.
• Better edge preservation.
• Reduced noise.
• Lower memory consumption compared to traditional methods.

---

🖼️ Sample Outputs

Input	FSRCNN	FSRCNN + AD	FSRCNN + AD + CG
Low Resolution	Enhanced	Refined	Final Output

---

🔮 Future Work

• Real-time image super-resolution.
• Deployment on mobile and edge devices.
• Integration with transformer-based architectures.
• Video super-resolution extension.

---

📚 Applications

• Medical Imaging
• Surveillance Systems
• Satellite Imaging
• Mobile Photography
• Remote Sensing

---

👨‍💻 Author

Venkat Raghava C G
M.Sc. Data Science
Vellore Institute of Technology, Chennai

📧 Email: raghavavenkat49@gmail.com

---

⭐ If you found this project useful, please consider giving it a star!