• This is the official repository of the paper "Revisiting Convolutional Sparse Coding for Image Denoising: From a Multi-scale Perspective" from IEEE Signal Processing Letters 2022. [Paper Link]

• Python >= 3.5
• PyTorch == 1.7.1 is recommended
• opencv-python = =3.4.9.31
• tqdm
• scikit-image == 0.15.0
• scipy == 1.3.1
• Matlab

The training data is from the [BSD500 dataset]. The test data are from [BSD500 dataset], [Urban100 dataset],[Kodak24 dataset],[McMaster dataset] Or you can download the datasets from our [Train Data], [Test Data].

1. Clone this repository:

   git clone https://github.com/JingyiXu404/MCSCNet.git
   

step 2 and step 3 can be ignored if you only use BSD68\Urban100\Kodak24\McMaster datasets as testsets (Download from our [Google Drive Link])

2. Generate .npy test datasets from original test data (follow the steps in dataset/test_data).

3. Place the high quality test images in dataset/test_data/your_folder. For example, dataset/test_data/gt_BSD68 .

   dataset 
   └── test_data
       ├── gt_BSD68
       └── gt_Urban100
       └── other test datasets
   

4. Run the following command for single image denoising task with different noise_levels and different datasets:

   python test.py
   

Modify variables dataset (line 42 in test.py) and noise_level (line 43 in test.py) to test with different datasets (BSD68/Urban100/Kodak24/McMaster) and different noise_levels (10/15/25/30/50/70/75/90)

5. Finally, you can find the Denoised results in ./test_results. Our results in the paper can be downloaded from [Google Drive Link]

If you find our work useful in your research or publication, please cite our work:

@article{xu2022revisiting,
  title={Revisiting Convolutional Sparse Coding for Image Denoising: From a Multi-scale Perspective},
  author={Xu, Jingyi and Deng, Xin and Xu, Mai},
  journal={IEEE Signal Processing Letters},
  year={2022},
  publisher={IEEE}
}

If you have any question about our work or code, please email [email protected] .