terryoo / mzsr

Meta-Transfer Learning for Zero-Shot Super-Resolution (CVPR, 2020)

MZSR

Meta-Transfer Learning for Zero-Shot Super-Resolution (Accepted for CVPR 2020)

Jae Woong Soh, Sunwoo Cho, and Nam Ik Cho

[Arxiv]

Environments

• Ubuntu 18.04
• Tensorflow 1.8
• CUDA 9.0 & cuDNN 7.1
• Python 3.6

Abstract

Convolutional neural networks (CNNs) have shown dramatic improvements in single image super-resolution (SISR) by using large-scale external samples. Despite their remarkable performance based on the external dataset, they cannot exploit internal information within a specific image. Another problem is that they are applicable only to the specific condition of data that they are supervised. For instance, the low-resolution (LR) image should be a "bicubic" downsampled noise-free image from a high-resolution (HR) one. To address both issues, zero-shot super-resolution (ZSSR) has been proposed for flexible internal learning. However, they require thousands of gradient updates, i.e., long inference time. In this paper, we present Meta-Transfer Learning for Zero-Shot Super-Resolution (MZSR), which leverages ZSSR. Precisely, it is based on finding a generic initial parameter that is suitable for internal learning. Thus, we can exploit both external and internal information, where one single gradient update can yield quite considerable results. (See Figure 1). With our method, the network can quickly adapt to a given image condition. In this respect, our method can be applied to a large spectrum of image conditions within a fast adaptation process.

Related Work

Super-Resolution for Various Kernels

[ZSSR (CVPR 2018)] "Zero-Shot" Super-Resolution Using Deep Internal Learning Link

[IKC (CVPR 2019)] Blind Super-Resolution With Iterative Kernel Correction Link

Optimization-based Meta-Learning

[MAML (ICML 2017)] Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks Link

[MAML++ (ICLR 2019)] How to train your MAML Link

Brief Description of Our Proposed Method

Illustration of the Overall Scheme

During meta-transfer learning, the external dataset is used, where internal learning is done during meta-test time. From random initial \theta_0, large-scale dataset DIV2K with “bicubic” degradation is exploited to obtain \theta_T. Then, meta-transfer learning learns a good representation \theta_M for super-resolution tasks with diverse blur kernel scenarios. In the meta-test phase, self-supervision within a test image is exploited to train the model with corresponding blur kernel.

Algorithms

  

Left: The algorithm of Meta-Transfer Learning & Right: The algorithm of Meta-Test.

Experimental Results

Results on various kernel environments (X2)

The results are evaluated with the average PSNR (dB) and SSIM on Y channel of YCbCr colorspace. Red color denotes the best results and blue denotes the second best. The number between parantheses of our methods (MZSR) denote the number of gradient updates.

Results on scaling factor (X4)

We will later provide the test input data

Visualized Results

Brief explanation of contents

├── GT: Ground-truth images
├── Input: Input LR images
├── Model: Pre-trained models are included
    ├──> Directx2: Model for direct subsampling (x2)
    ├──> Multi-scale: Multi-scale model
    ├──> Bicubicx2: Model for bicubic subsampling (x2)
    └──> Directx4: Model for direct subsampling (x4)
└── results: Output results are going to be saved here.

Rest codes are for the test of MZSR.

Guidelines for Codes

Requisites should be installed beforehand.

Clone this repo.

git clone http://github.com/JWSoh/MZSR.git
cd MZSR/

Test

Ready for the input data (low-resolution) and corresponding kernel (kernel.mat file.)

[Options]

python main.py --gpu [GPU_number] --inputpath [LR path] --gtpath [HR path] --savepath [SR path]  --kernelpath [kernel.mat path] --model [0/1/2/3] --num [1/10]

--gpu: If you have more than one gpu in your computer, the number designates the index of GPU which is going to be used. [Default 0]
--inputpath: Path of input images [Default: Input/g20/Set5/]
--gtpath: Path of reference images. [Default: GT/Set5/]
--savepath: Path for the output images. [Default: results/Set5]
--kernelpath: Path of the kernel.mat file. [Default: Input/g20/kernel.mat]
--model: [0/1/2/3]
    -> 0: Direct x2
    -> 1: Multi-scale
    -> 2: Bicubic x2
    -> 3: Direct x4
--num: [1/10] The number of adaptation (gradient updates). [Default 1]

You may change other minor options in "test.py." Line 9 to line 17.

The minor options are shown below.

self.save_results=True		-> Whether to save results or not.
self.display_iter = 1		-> The interval of information display.
self.noise_level = 0.0		-> You may sometimes add small noise for real-world images.
self.back_projection=False	-> You may also apply back projection algorithm for better results.
self.back_projection_iters=4	-> The number of iteration of back projection.

An example of test codes

python main.py --gpu 0 --inputpath Input/g20/Set5/ --gtpath GT/Set5/ --savepath results/Set5 --kernelpath Input/g20/kernel.mat --model 0 --num 1

Citation

Will be updated soon.

Acknowledgement

Our work and implementations are inspired by and based on ZSSR [site] and MAML [site].