AKDT: Adaptive Kernel Dilation Transformer for Effective Image Denoising

🆕 Updates

• 14.04.2026 📢🥳 Check out our new CVPR 2026 paper: Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex: 📄Project Page, 📎Paper, ⭐Code.
• 28.07.2025 ✨ Check out our new multimodal framework: ModalFormer: Multimodal Transformer for Low-Light Image Enhancement! Paper and HF Demo coming soon!
• 20.04.2025 🏆 We use AKDT to participate in the CVPR Workshop NTIRE 2025 Image Denoising Challenge. Check out the Challenge Report.
• 23.03.2025 🔗 Repository updated with link to paper PDF.
• 04.12.2024 🎉 Paper has been accepted at VISAPP 2025. To be published.

🧪 Experiment

⚙️ 1. Create Environment

• Make Conda Environment

conda create -n AKDT python=3.7
conda activate AKDT

• Install Dependencies

conda install pytorch=1.8 torchvision cudatoolkit=10.2 -c pytorch
pip install matplotlib scikit-learn scikit-image opencv-python yacs joblib natsort h5py tqdm
pip install einops gdown addict future lmdb numpy pyyaml requests scipy tb-nightly yapf lpips

• Install basicsr

python setup.py develop --no_cuda_ext

📁 2. Prepare Datasets

Download the datasets and place them as specified in the ./Denoising/Datasets/README.md

SIDD_train - Google Drive

SIDD_val - Google Drive

SIDD_test - Google Drive

BSD400 - Google Drive

DIV2K - Google Drive

WaterlooED - Google Drive

gaussian_test - Google Drive

🧫 3. Test

Pre-trained weights available at Google Drive. Place the pre-trained weights into the ./Denoising/pretrained_models/ directory.

• Real Image Denoising evaluation

# To obtain denoised results
python test_real_denoising_sidd.py --save_images
# Compute PSNR
python eval_sidd.py

• Color Gaussian Image Denoising evaluation

# To obtain denoised results
python test_gaussian_color_denoising.py --model_type blind --sigmas 15,25,50
# Compute PSNR
python evaluate_gaussian_color_denoising.py --model_type blind --sigmas 15,25,50

Note: --weights argument can be used to specify paths to different weights.

📉 4. Compute Complexity

You can test the model complexity (FLOPS/MACs/Params) using the following command:

python ./basicsr/models/archs/macs.py

🏋️ 5. Train

• Generate training image patches:

# Gaussian color image denoising
python generate_patches_dfwb.py 
# Real image denoising
python generate_patches_sidd.py 

• Train AKDT on Color Gaussian Image Denoising:

./train.sh Denoising/Options/GaussianColorDenoising_AKDT.yml

• Train AKDT on Real Image Denoising:

./train.sh Denoising/Options/GaussianColorDenoising_AKDT.yml

📚 Citation

@inproceedings{brateanu2025akdt,
  author    = {Alexandru Brateanu and Raul Balmez and Adrian Avram and Ciprian Orhei},
  title     = {AKDT: Adaptive Kernel Dilation Transformer for Effective Image Denoising},
  booktitle = {Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 3: VISAPP},
  year      = {2025},
  pages     = {418--425},
  isbn      = {978-989-758-728-3},
  issn      = {2184-4321}
}