https://github.com/lusinlu/gradient-variance-loss

Code of the ICASSP 2022 paper "Gradient Variance Loss for Structure Enhanced Super-Resolution"
https://github.com/lusinlu/gradient-variance-loss

deep-learning loss loss-functions machine-learning model-optimization super-resolution training upscaling

Last synced: 4 months ago
JSON representation

Code of the ICASSP 2022 paper "Gradient Variance Loss for Structure Enhanced Super-Resolution"

• Host: GitHub
• URL: https://github.com/lusinlu/gradient-variance-loss
• Owner: lusinlu
• Created: 2021-08-25T08:14:07.000Z (about 4 years ago)
• Default Branch: main
• Last Pushed: 2022-02-27T16:05:14.000Z (over 3 years ago)
• Last Synced: 2025-04-14T07:47:08.874Z (6 months ago)
• Topics: deep-learning, loss, loss-functions, machine-learning, model-optimization, super-resolution, training, upscaling
• Language: Python
• Homepage:
• Size: 204 KB
• Stars: 31
• Watchers: 1
• Forks: 0
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Gradient Variance Loss

[![report](https://img.shields.io/badge/arxiv-report-red)](https://arxiv.org/abs/2202.00997)

[ICASSP 2022] Official implementation of the Gradient Variance loss presented in the paper [paper](https://arxiv.org/abs/2202.00997) "Gradient Variance Loss for Structure-Enhanced Image Super-Resolution".

## Requirements

for installing required packages run

` pip install -r requirements.txt`

## Usage

To train the VDSR model with the gradient variance loss run the following command

` python train.py --dataroot [path to DIV2K dataset] --cuda`

## Introduction

"Gradient Variance Loss for Structure-Enhanced Image Super-Resolution"

By Lusine Abrahamyan, Anh Minh Truong, Wilfried Philips and Nikos Deligiannis.

### Approach

We observe that gradient maps of images generated

by the models trained with the L1/L2 losses have significantly lower variance than the gradient maps of the original

high-resolution images. 

In this work, we introduce a structure-enhancing loss

function, coined Gradient Variance (GV) loss, to minimize the difference between the variances of predicted and original gradient maps and generate

textures with perceptual-pleasant details.

### Performance

Public benchmark test results and DIV2K validation results (PSNR(dB) / SSIM).



## Citation

If you find the code useful for your research, please consider citing our works

```

@article{abrahamyangvloss,

  title={Gradient Variance Loss for Structure-Enhanced Image Super-Resolution},

  author={Lusine, Abrahamyan and  Anh Minh, Truong and  Wilfried, Philips and Nikos, Deligiannis},

  journal={Proceedings of the International Conference on Acoustics, Speech, and Signal Processing (ICASSP)},

  publisher = {IEEE},

  year={2022}

}

```

## Acknowledgement

Codes for the VDSR model are from [pytorch-vdsr](https://github.com/twtygqyy/pytorch-vdsr).