Experiments on ESRGAN models with attention modules

This repository contains models based on the enhanced deep residual networks for single image super-resolution
(EDSR), with the addition of various attention modules to improve the upscaling results.

SRGAN was one of the first generative adversarial model trained for single image super-resolution up to an upscaling factor of 4. Based on that,
enhanced SRGAN (ESRGAN) improved the upscaling performance by modifying the overall architecture. In the generator model, batch normalization is removed and residual-in-residual dense blocks (RRDB) are used to increase depth and connectivity of the network. Furthermore, a relativistic average discriminator is used, which predicts predicts a realsitic or fake batch relative to the ground truth data distribution.

DIV2K is a widely used dataset for training super-resolution models. It consists of 1000 high-definition RGB images
of with a 2K resolution and a large diversity of content, split into a batch of 800 training images and 100 images each for validation and testing.
For testing purposes, the widely used super-resolution datasets such as Set5, Set14 and Urban100 are used.

DIV2K can be downloaded using this link: https://data.vision.ee.ethz.ch/cvl/DIV2K/.

The metrics used for evaluating model performance and image reconstruction quality in general are
the pixelwise signal-to-noise ratio PSNR and the structural similarity index metric SSIM. \

This table shows the results for the different dicriminator models. The discriminator architectures marked with PA are
enhanced with a pyramid attention mechanism.