This repository implements the code of Deep Model-Based Super-Resolution with Non-uniform Blur

Train

To train the code please first download COCO dataset available at: https://cocodataset.org.

python main_train.py -opt options/train_nimbusr.json

Test

Pre-trained model is available at: model_zoo/DMBSR.pth

Our blur kernels are available for download here. They need to be added in the folder |-kernels

See test_model.ipynb to test the model on COCO dataset.

Results

We achieve state-of-the-art results in super-resolution in the presence of spatially-varying blur. Here are some of the results we obtained. Feel free to test on your own sample using the testing notebook.

LR	SwinIR	BlindSR	USRNet	Ours	HR

LR	SwinIR	BlindSR	USRNet	Ours

Real-world images

For this section, we used the code provided by https://github.com/GuillermoCarbajal/NonUniformBlurKernelEstimation to estimate the kernel and we combine their kernel estimation to our super-resolution model. We also use the dataset provided by "Laurent D’Andrès, Jordi Salvador, Axel Kochale, and Sabine Süsstrunk. Non-parametric blur map regression for depth of field extension".

Defocus x2 super-resolution

LR	SwinIR	BlindSR	Ours

Deblurring

LR	DMPHN	RealBlur	MPRNet	Ours

Acknowledgement

The codes use KAIR as base. Please also follow their licenses. I would like to thank them for the amazing repository.

Citation

If you use our work, please cite us with the following:

@InProceedings{laroche2023dmbsr,
  title = {Deep Model-Based Super-Resolution with Non-Uniform Blur},
  author = {Laroche, Charles and Almansa, Andrés and Tassano, Matias},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)}
  year = {2023}
}