Learning a Single Convolutional Super-Resolution Network for Multiple Degradations

PyTorch training and testing code - 18/12/2019

I recommend to use the PyTorch code for training and testing. The model parameters of MatConvnet and PyTorch are same.

• main_train_srmd.py
• main_test_srmd.py

Abstract

Recent years have witnessed the unprecedented success of deep convolutional neural networks (CNNs) in single image super-resolution (SISR). However, existing CNN-based SISR methods mostly assume that a low-resolution (LR) image is bicubicly downsampled from a high-resolution (HR) image, thus inevitably giving rise to poor performance when the true degradation does not follow this assumption. Moreover, they lack scalability in learning a single model to deal with multiple degradations. To address these issues, we propose a dimensionality stretching strategy that enables a single convolutional super-resolution network to take two key factors of the SISR degradation process, i.e., blur kernel and noise level, as input. Consequently, the proposed super-resolver can handle multiple and even spatially variant degradations, which significantly improves the practicability. Extensive experimental results on synthetic and real LR images show that the proposed convolutional super-resolution network not only can produce favorable results on multiple degradations but also is computationally efficient, providing a highly effective and scalable solution to practical SISR applications.

Training Codes

SRMD_MatConvNet

Network Architecture

In contrast to other CNN-based SISR methods which only take the LR image as input and lack scalability to handle other degradations, the proposed network takes the concatenated LR image and degradation maps as input, thus allowing a single model to manipulate multiple and even spatially variant degradations.

NTIRE 2018 Challenge on Image Super-Resolution 知乎

Track 3: Realistic Difficult Adverse Conditions

Test_Track_3_SRMD_non_blind.m is the testing code for Track 3 of NTIRE 2018 challenge on image super-resolution.

PSNR/SSIM Result of SRMD

Visual Result of SRMD

The image size of the LR image is 510x405, the scale factor is 4.

SISR Results

Bicubic Degradation

General Degradation

The left is the LR image with isotropic Gaussian blur and noise level 50, the right is the super-resolved image by SRMD with scale factor 2.

The left is the LR image with isotropic Gaussian blur and noise level 5, the right is the super-resolved image by SRMD with scale factor 3.

The left is the LR image with anisotropic Gaussian blur, the right is the super-resolved image by SRMD with scale factor 4.

Real Image SR

SRMD models

• Bicubic Degradation

  Demo_bicubic_degradation_SRMD.m is the testing demo of SRMD for the widely-used bicubic degradation. Demo_bicubic_degradation_SRMDNF.m is the testing demo of SRMDNF for the widely-used bicubic degradation.

• General Degradation

  Demo_genearal_degradation_SRMD.m is the testing demo of SRMD for the general degradation. Demo_genearal_degradation_SRMDNF.m is the testing demo of SRMDNF for the general degradation.

• Real Application

  Demo_real_application_SRMD.m is the testing demo of SRMD for real image SR.

• Others

  Demo_degradation_direct_SRMD.m is the testing demo of SRMD for another widely-used degradation with direct downsampler.

  Demo_degradation_direct_SRMDNF.m is the testing demo of SRMD for another widely-used degradation with direct downsampler.

  A special case of SRMD when scale factor is 1.

  Demo_denoising_and_deblurring_gray_SRMD.m is the testing demo of SRMD for grayscale image denoising and deblurring.

  Demo_denoising_and_deblurring_color_SRMD.m is the testing demo of SRMD for for color image denoising and deblurring.

  Demo_real_application_denoising_and_deblurring_gray.m is the testing demo of SRMD for real grayscale image denoising and deblurring.

  Demo_real_application_denoising_and_deblurring_color.m is the testing demo of SRMD for real color image denoising and deblurring.

Requirements and Dependencies

• MATLAB R2015b

• Cuda-8.0 & cuDNN v-5.1

• MatConvNet

  or

• MATLAB R2015b

  Use `res = vl_srmd_matlab(net, input)` instead.

Citation

@inproceedings{zhang2018learning,
  title={Learning a single convolutional super-resolution network for multiple degradations},
  author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  pages={3262-3271},
  year={2018}
}