srgan

This repository is experimental implementation of "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network" ( SRGAN )

https://arxiv.org/abs/1609.04802v5

data

The model was trained with 800 DIV2k dataset.

You can download it from https://data.vision.ee.ethz.ch/cvl/DIV2K/

The model was tested with set5, and set14 dataset

About SRGAN

SRGAN is a deep learning model used for image super-resolution, which is generating high-resolution images from their low-resolution counterparts. Previous deep learning-based methods for super-resolution typically focused on minimizing pixel-wise loss functions, such as mean squared error (MSE). However, this approach can lead to overly smooth images with poor perceptual quality since it encourages the model to find a pixel-wise average of the possible solutions. In contrast, SRGAN focuses on generating images with better perceptual quality by using Generative Adversarial Network and perceptual loss.

Generative Adversarial Network (GAN)

Generative Adversarial Network (GAN) is model consisted of 2 networks, Generator Network and Discriminator Network. In SRGAN, The generator network produces high-resolution images from low-resolution input, while the discriminator network distinguishes generated images and original high-resolution images.This results in generator creating images with better perceptual quality

Perceptual loss

The loss function used in SRGAN is called perceptual loss, and it consists of 2 parts

The Adversarial loss is from GAN structure

The content loss is defined as the Euclidean distance between the VGG19 feature representations of the generated output and the original high-resolution image. The use of content loss forces the network to be trained to produce images with better perceptual quality. This is because the deep VGG-19 feature maps extract high-level content information from the image, which helps the SRGAN network to focus on restoring a realistic image rather than relying solely on pixel-wise similarity.

Results

Original HR image

x4 Low-resolution image

SRResNet

SRGAN

Original HR image

x4 Low-resolution image

SRResNet

SRGAN