Hao Li, Zhijing Yang, Xiaobin Hong, Ziying Zhao, Junyang Chen, Yukai Shi, Jinshan Pan

Guangdong University of Technology, South China University of Technology, Sun Yat-sen University, Nanjing University of Science and Technology

• Released the training codes and the test results.

Real-world image denoising is a practical image restoration problem that aims to obtain clean images from in-the-wild noisy input. Recently, Vision Transformer (ViT) exhibits a strong ability to capture long-range dependencies and many researchers attempt to apply ViT to image denoising tasks. However, real-world image is an isolated frame that makes the ViT build the long-range dependencies on the internal patches, which divides images into patches and disarranges the noise pattern and gradient continuity. In this article, we propose to resolve this issue by using a continuous Wavelet Sliding-Transformer that builds frequency correspondence under real-world scenes, called DnSwin. Specifically, we first extract the bottom features from noisy input images by using a CNN encoder. The key to DnSwin is to separate high-frequency and low-frequency information from the features and build frequency dependencies. To this end, we propose Wavelet Sliding-Window Transformer that utilizes discrete wavelet transform, self-attention and inverse discrete wavelet transform to extract deep features. Finally, we reconstruct the deep features into denoised images using a CNN decoder. Both quantitative and qualitative evaluations on real-world denoising benchmarks demonstrate that the proposed DnSwin performs favorably against the state-of-the-art methods.

If you find this work useful for your research, please cite:

@artical{Li2022dnswin,
  title={DnSwin: Toward Real-World Denoising via Continuous Wavelet Sliding-Transformer},
  author={Li, Hao and Yang, Zhijing and Hong, Xiaobin and Zhao, Ziying and Chen, Junyang and Shi, Yukai and Pan, Jinshan},
  journal={Knowledge-Based Systems},
  year={2022}
}

You can download the test results (.mat) of our DnSwin in [Google Drive] and evaluate the metrics in the public online page SIDD Benchmark (Method: DnWSTrans) and DND Benchmark (Method: DnSwin).

Download the training datasets from GoogleDrive.

Extract the files to data folder as follow:

~/
  data/
    SIDD_train/
      ... (scene id)
    SIDD_valid/
      ... (id)
    Syn_train/
      ... (id)
    DND/
      images_srgb/
        ... (mat files)
      ... (mat files)

The code to generate a synthetic dataset is provided by GMSNet .

The code you can find in utils/syn.

Copy the template code to build your own model:

~/
  model/
    DnSwin.py
    template.py
    ... (your model)

Train your own model:

python train.py --model ... (model name)

Evaluate the trained model (--ensemble for higher score):

python submit_dnd.py --model ... (model name) --ensemble

The results are in result/submit_dnd/bundled

Please contact me if there is any question (Hao Li: [email protected]).