Super-Resolution Image Enhancement

Overview

This project implements Single Image Super-Resolution (SISR) using deep learning to reconstruct high-resolution images from their low-resolution counterparts. We compare three CNN architectures and explore different loss functions to achieve optimal image quality.

Key Features

• 3 Model Architectures: SimpleCNN (baseline), MobileNetV2 (transfer learning), ResNet18-UNet (advanced)
• Loss Function Comparison: L1, L2 (MSE), and Perceptual Loss (VGG16-based)
• Data Augmentation: Flips, rotations, brightness/contrast adjustments
• Comprehensive Evaluation: PSNR and SSIM metrics on test set

Results

• Best Model: ResNet18-UNet achieves ~33-35 dB PSNR and 0.91-0.94 SSIM
• Best Loss: L2 Loss performs slightly better (PSNR=32.96 dB, SSIM=0.9648)
• Transfer Learning: Significantly outperforms training from scratch with limited data

Dataset

Source: Image Super Resolution from Unsplash

• 200 images (256×256 RGB)
• 4× upscaling factor
• 70% train / 15% val / 15% test split

Project Structure

project/
├── project3_report.ipynb    # Complete technical report with code
├── requirements.txt         # Python dependencies
├── dataset/                 # High-res and low-res images (hidden)
└── checkpoints/            # Saved models and training histories

Quick Start

# Install dependencies
pip install -r requirements.txt

# Run the notebook
jupyter notebook project3_report.ipynb

Technologies

• PyTorch 2.0+
• TorchVision (Pre-trained models)
• NumPy, Pandas, Matplotlib
• PSNR/SSIM metrics for evaluation