WARNING: This repository is still a work in progress. The code may break at any time.
This repository contains the code for implementation of the approach outlined in the paper Single image super-resolution using Gaussian process regression by He & Siu. As the name suggests, the approach uses Gaussian Process Regression to perform super-resolution on a single image.
The code is written in Python 3.10.12 and uses the GPytorch library for Gaussian Process Regression. The code is tested on MacOS 13.4.1.
Test pictures are taken from the Set14 dataset, which was downloaded from here. These can be found in the Set14 folder. I have also generated smaller versions (2x) of the images, which can be found in the Set14_smaller folder.
Currently, the following kernels are implemented and tested:
- RBF (Squared Exponential)
- Matern 3/2
- Matern 5/2
- Exponential
In general, the Matern 3/2 and Matern 5/2 kernels perform better than the other options.
The main entry point for the code is the main.py file. The file contains the code for generating the high resolution image from the low resolution image. There are many options for user customization, which I have tried to document in the file itself.
Additionally, two helpful shell scripts for analyzing the results (i.e. calulating SSIM and PSNR values) are provided. Note that these scripts require ImageMagick to be installed. The scripts are:
comparison_methods.sh: This script calculates the average SSIM and PSNR for each of the methods (bicubic interpolation, different kernels in Gaussian Process Regression, etc.).comparison_images.sh: This script calculates the SSIM and PSNR for each of the images. This allows us to see which images are better suited for the different methods.