SRGAN for Image Super-Resolution on DIV2K Dataset

This repository contains a Jupyter Notebook implementation of the Super-Resolution Generative Adversarial Network (SRGAN), a deep learning model designed for image super-resolution. The model is trained and evaluated on the DIV2K dataset, which is widely used for benchmarking super-resolution algorithms.

Features

• SRGAN Architecture: Combines a generator (for high-resolution image generation) and a discriminator (to distinguish real from generated images) in an adversarial setup.
• Perceptual Loss: Uses a combination of content loss (based on VGG features) and adversarial loss for sharper, realistic outputs.
• DIV2K Dataset: High-quality dataset of 2K resolution images for super-resolution tasks.
• Training and Evaluation: Demonstrates how to preprocess data, train SRGAN, and evaluate the results.

Installation

1. Clone the repository and navigate to the project directory.
2. Install the required dependencies using:

   pip install -r requirements.txt

Dataset

Download the DIV2K dataset from the official website and place it in the data/ directory. Update the dataset path in the notebook as needed.

Usage

1. Open the notebook file srgan-on-div2k.ipynb in Jupyter Notebook or JupyterLab.
2. Follow the steps in the notebook to:
   • Preprocess the DIV2K dataset
   • Define the SRGAN architecture
   • Train the SRGAN model
   • Evaluate the model on test images

Results

The notebook provides qualitative and quantitative results demonstrating the improvement in image resolution. Example outputs include side-by-side comparisons of low-resolution inputs, SRGAN outputs, and ground truth high-resolution images.

References

• Ledig et al., Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
• DIV2K Dataset