By Zheming Zuo¹, Xinyu Chen², Han Xu^3,4, Jie Li⁵, Wenjuan Liao⁶, Zhi-Xin Yang², and Shizheng Wang^4
1 Department of Computer Science, Durham University, Durham DH1 3LE, UK
² Department of Electromechanical Engineering, University of Macau, Macau 999078, China
³ University of Chinese Academy of Sciences, Beijing 100049, China
⁴ Institute of Microelectronics, Chinese Academy of Sciences, Beijing 100029, China
⁵ School of Computing, Engineering & Digital Technologies, Teesside University, Middlesbrough TS3 6DR, UK
⁶ College of Engineering and Computer Science, Australian National University, Canberra ACT2600, Australia

This is an official implementation of our adaptIve Dual sElf-Attention Network (IDEA-Net).

IDEA-Net is an unsupervised single image denoiser, which performs on superiorly on AWGN and real-world image noises with a single end-to-end deep neural network, and contributes to the downstream task of face detection in low-light conditions. Our work is inspired by Unsupervised Attention-guided Image-to-Image Translation and Self2Self With Dropout: Learning Self-Supervised Denoising From Single Image.

For more details, please refer our

• [paper]
• [supp]

1. Preparation
2. Installation
3. Run
4. Performance

Clone the github repository.

  git https://github.com/zhemingzuo/IDEA-Net --recursive
  cd IDEA-Net

Set up a conda environmentwith all dependencies as follows:

  conda env create -f environment.yml
  source activate ideanet_wacv_2022

Please note our implementation is only tested under Ubuntu environment with Nvidia GPUs and CUDA installed.

Run our IDEA-Net for image denosing via

  python main.py  --to_train=1 --log_dir=./output/exp_01 --config_filename=./configs/exp_01.json

1. Removing AWGN Image Noise

Comparisons of AWGN denoising results in terms of PSNR on the (C)BSD68 dataset with valued as 25 and 50. denotes the selected image region for comparison and indicates the dual self-attention region drawn by IDEA-Net. Best viewed in zoomed mode.

2. Removing Real-World Image Noise

Comparisons of real-world image noise removal results with respect to PSNR on the PolyU dataset. denotes the selected image region for comparison and indicates the dual self-attention region drawn by the proposed IDEA-Net. Best viewed in zoomed mode..

3. Downstream task on dark face detection

Performance comparisons of real-world dark/noisy face detection on the DARK FACE dataset. Light-Enhanced Noisy Image (LENI) is yielded by MSRCR. Detection results are generated by a RetinaNet that pre-trained on the WIDER FACE dataset. and respectively represents the correct and erroneous detections. indicates the dual self-attention region drawn by IDEA-Net. Best viewed in zoomed mode.

If you find IDEA-Net useful in your research, please consider citing:

@InProceedings{Zuo_IdeaNet_2022_WACV,
    author    = {Z. Zuo and X. Chen and H. Xu and J. Li and W. Liao and Z.-X. Yang and S. Wang},
    title     = {IDEA-Net: Adaptive Dual Self-Attention Network for Single Image Denoising},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops},
    month     = {January},
    year      = {2022},
    pages     = {739-748}
}