ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks¶

Star | Fork | Issue

Pipeine for Image Super-Resolution task that based on a frequently cited paper, ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks (Wang Xintao et al.), published in 2018.

In few words, image super-resolution (SR) techniques reconstruct a higher-resolution (HR) image or sequence from the observed lower-resolution (LR) images, e.g. upscaling of 720p image into 1080p.

One of the common approaches to solving this task is to use deep convolutional neural networks capable of recovering HR images from LR ones. And ESRGAN (Enhanced SRGAN) is one of them. Key points of ESRGAN:

SRResNet-based architecture with residual-in-residual blocks;
Mixture of context, perceptual, and adversarial losses. Context and perceptual losses are used for proper image upscaling, while adversarial loss pushes neural network to the natural image manifold using a discriminator network that is trained to differentiate between the super-resolved images and original photo-realistic images.

Technologies¶

Catalyst as pipeline runner for deep learning tasks. This new and rapidly developing library can significantly reduce the amount of boilerplate code. If you are familiar with the TensorFlow ecosystem, you can think of Catalyst as Keras for PyTorch. This framework is integrated with logging systems such as the well-known TensorBoard.
Pytorch, torchvision, and PIQ as main frameworks for deep learning.
Albumentations for data preprocessing.

Quick Start¶

Setup environment¶

esrgan requires python >= 3.8. The requirements.txt file can be used to install the necessary packages.

git clone  https://github.com/leverxgroup/esrgan.git
pip install ./esrgan

Run an experiment¶

# step 1 - supervised training of the model
catalyst-dl run -C esrgan/experiment/config_supervised.yml --benchmark

# step 2 - use weights from step 1 to train model using GAN approach
catalyst-dl run -C esrgan/experiment/config_gan.yml --benchmark