Skip to content

Latest commit

 

History

History
98 lines (79 loc) · 3.77 KB

README.md

File metadata and controls

98 lines (79 loc) · 3.77 KB
Raw

BLIND USRNET

A project in Statistical Image Processing. Extension of Deep unfolding network for image super-resolution for Blind Super Resolution via Kernel Estimation

In this project we aim to extend USRNet [1] for Blind Super Resolution. We added a step of blind kernel estimation [2] and noise STD estimation [3] algorithm. These additions show SOTA performance in Blind-SR

Default

Getting Started

Clone the Repo:

git clone https://github.com/geopi1/Improved_USRNet.git

Datasets

Download the Datasets (a bicubically downscaled version of DIV2K):

Download DIV2K
This link will download 2 folders:

  • DIV2K_LR - the low resolution input
  • DIV2K_HR_ds2 - Ground truth high resolution images

Prerequisites

Tested and ran on:

  • UBUNTU 18.04
  • RTX 2080
  • Nvidia driver 440.95.10
  • cuda 10.1.243
  • cudnn 7.6.5
  • pytorch 1.5
  • TF 1.14
    for additional package information please refer to requirements.txt or env.yml
  1. Setup conda 
    conda env create -f env.yml
    This will create a working environment named Blind_USRNet
  2. Setup can also be performed with pip (virtual env) via the requirements.txt file 
    python -m venv Blind_USRNet
pip install -r requirements.txt
  3. There are several examples provided with the code in ./input_images
  4. All other images to be tested should be placed in ./input_images

Running the Code

Code

python main.py

This will output 4 types of images to ./results:

  • Default USRNet settings
  • Default USRNet settings + estimated Noise STD
  • Kernel estimated USRNet
  • Kernel estimated USRNet + estimated Noise STD

Additionally, for each estimated image a side-by-side image with the LR version is saved with the degredation kernel (default or estimated)

Numerical Evaluation

To calculate the numerical results on the whole dataset run:

python utils/get_results.py

This will calculate the mean PSNR and SSIM on the SR<->HR on all the results

Numerical Results

Metric Default Default + Noise Est. Kernel Est. Kernel Est. + Noise Est.
PSNR 22.35 dB 22.30 (-0.05) dB 26.31 (+3.96) dB 26.29 (+3.94) dB
SSIM 0.756 0.755 (-0.001) 0.797 (+0.041) 0.797 (+0.041)

Visual Results

On the left, the LR image and on the Right the reconstructed HR image.
The blur kernel of the LR image is at the top left corner.
Default:
Default
Ours:
Ours
Default:
Default
Ours:
Ours

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgments

[1] Zhang, Kai, Luc Van Gool, and Radu Timofte. "Deep unfolding network for image super-resolution." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.

[2] Bell-Kligler, Sefi, Assaf Shocher, and Michal Irani. "Blind super-resolution kernel estimation using an internal-gan." Advances in Neural Information Processing Systems. 2019.

[3] Chen G , Zhu F , Heng P A . "An Efficient Statistical Method for Image Noise Level Estimation" 2015 IEEE International Conference on Computer Vision (ICCV). IEEE Computer Society, 2015.