Awesome

COLA-Net: Collaborative Attention Network for Image Restoration

This repository is for COLA-Net introduced in the following paper: Chong Mou, Jian Zhang, Xiaopeng Fan, Hangfan Liu, and Ronggang Wang, "COLA-Net: Collaborative Attention Network for Image Restoration", (IEEE Transactions on Multimedia 2021)

[Paper] [arxiv]

Our Team

Requirements

The code is built based on RNAN.

• Python 3.6
• PyTorch == 1.1.0
• numpy
• skimage
• cv2

The training datasets are available at DIV2K and SIDD.

Contents

1. Introduction
2. Tasks
3. Citation
4. Acknowledgements

Introduction

In this paper we propose a model dubbed COLA-Net to exploit both local attention and non-local attention to restore image content in areas with complex textures and highly repetitive details, respectively. It is important to note that this combination is learnable and self-adaptive. To be concrete, for local attention operation, we apply local channel-wise attention on different scales to enlarge the size of receptive field of local operation, while for non-local attention operation, we develop a novel and robust patch-wise non-local attention model for constructing long-range dependence between image patches to restore every patch by aggregating useful information (self-similarity) from the whole image.

The pre-trained models are available at Google Drive and PKU Drive.

Proposed COLA-Net

1. The gloabal architecture of our proposed COLA-Net.
2. The details of our proposed patch-based non-local attention method.
3. The visualization of the collaborative attention mechanism.

Tasks

Gray-scale Image Denoising

Real Image Denoising

Image Compression Artifact Reduction

Citation

If you find the code helpful in your resarch or work, please cite the following papers.

@inproceedings{zhang2019rnan,
    title={Residual Non-local Attention Networks for Image Restoration},
    author={Zhang, Yulun and Li, Kunpeng and Li, Kai and Zhong, Bineng and Fu, Yun},
    booktitle={ICLR},
    year={2019}
}

@article{mou2021cola,
  title={COLA-Net: Collaborative Attention Network for Image Restoration},
  author={Chong, Mou and Jian, Zhang and Xiaopeng, Fan and Hangfan, Liu and Ronggang, Wang},
  journal={IEEE Transactions on Multimedia},
  year={2021}
}

Acknowledgements

This code is built on RNAN (PyTorch). We thank the authors for sharing their codes of RNAN.