Super Resolution by Neural Networks

Introduction

Implementation for the super resolution of both 2D and 3D images. Involved Neural Networks: EDSR, SR-Resnet Residual, SRGAN, WDSR, SRCNN.

Authors

Ying Da Wang
Yanfeng Li
Ryan Armstrong
Peyman Mostaghimi
...

(School of Minerals and Energy Resources Engineering, Univerisity of New South Wales)

Coming Soon:
Migration to TF2, and unpaired real LR to HR networks

Installation

Version 0.1
This software is compatible with windows, mac, and linux machines. It will install anaconda3, along with the necessary python packages in a containerised anaconda environment.

Before continuing with installation, you will need to download this repository, and a folder containing the models and testing images.

Downloading this repo can be done using the link above - click on the green "Code" button, and then click on "Download ZIP".

Once Downloaded, unzip the folder and download the model (see next sentence)

The folder containing the models and testing images is here: https://drive.google.com/file/d/13o3Vz65YlByJjw8zMvX0C11TXCvKGOXS/view?usp=sharing

Download and extract this into each folder "Windows, Mac, or Linux" - whichever you wish to use.

Linux and Mac

Open a terminal window at the directory where the file “installSR.sh” is located and type “bash installSR.sh”.

Step1:

Open a “Anaconda cmd Prompt” terminal window at the directory where the file “installSR.sh” is located and type “bash installSR.sh”.

$ bash installSR.sh

Step2:

Please input this command:

For Mac, please use

$ bash installSRMac.sh

For Linux, please use

$ bash installSRLinux.sh

Step3:

Always type “y” for yes if it is required.

Step4:

Please input this command to run the software:

For Mac, please use

$ bash runSRMac.sh

For Linux, please use

$ bash runSRLinux.sh

FOR ADVANCED USERS: if you already have anaconda3, or wish to install the packages yourself, a full list of conda packages used by this software are shown below:

conda install tensorflow=1.13.0 

conda install matplotlib 

conda install pillow 

conda install -c conda-forge gooey  

pip install tensorlayer==1.11 

pip install argparse 

Windows

For Windows system, we reccomand you to install via the conda promt.

If you do not have the anaconda3, you may download it via this link:
https://repo.anaconda.com/archive/Anaconda3-2019.03-Windows-x86_64.exe

For Linux system, click this link: https://repo.anaconda.com/archive/Anaconda3-2019.03-Linux-x86_64.sh

Then, please open a “Anaconda Prompt” terminal window shown in the below image.

The specific installation steps are as follow. Always type "y" for "yes".

conda update -n base -c defaults conda

conda create --name srRockEnv2 python=3.6

conda activate srRockEnv2

For linux system, you may prefer to type "conda install tensorflow-gpu=1.13.0":

conda install tensorflow-gpu=1.13.0

For Mac system, "conda install tensorflow=1.13.0" instead:

conda install tensorflow=1.13.0

Then, please type these lines:

conda install matplotlib

conda install pillow

conda install -c conda-forge gooey

pip install tensorlayer==1.11

pip install argparse

##Running Guideline##

A desktop shortcut icon should have been created as part of the installation. Please double click it. In case we can't get that to work:
Open a terminal window at the directory where the file “runProgram.sh” is located and type “bash runProgram.sh”.

Input Format There are 5 different formats of input image: .png, .jpg, .mat, .nc, .tif.

2D Images All input 5 input formats listed above are acceptable for 2D images resolution.

3D Images The format of .mat, .nc and .tif are acceptable for 3D image super resolution.

parameter	usage
Input images	Folder which containing input images, named “srtestfolder”
Input format	The format of input image
Scale factor	Up sampling factor
Output format	The format of output image
Bit depth	Bit depth of your input images
Image dimension	2D or 3D dimension of input images
Use CPU	Force the network to use the CPU as default, if no compatible GPU is detected
Use GAN	Force the network to use the GAN as default

Output format and Visualisation

The output images are stored in “srtestfolder\srOutputs”.

For 2D images, output is slice-by-slice, and readable my standard image reading software.

For 3D images, we recommend ImageJ: https://imagej.net/Fiji/Downloads

Examples

2D Images

1. choose the folder containing the input images on your computer;

2. choose the up-sampling scale factor you want, such as 4, 16 or 64;

3. when you choose to resolution 2D images, it may not be necessary to use 3D patches;

4. there are only two types of 2D output formats, “.png” and “.jpg”, could be used;

5. bit depth: uint8 means unsigned 8-bits integer;

6. choose 2D as image dimension;

7. if you choose “yes” as “USE CPU”, the software will force the network to use CPU rather than GPU. And CPU will as default if you do not choose;

8. there are two types of neural network could be provided, such as CNN and GAN. And the CNN will as default if you do not choose;

9. determine to use checkpoints for training 2D images or not.

10. Start to resolution your images!

3D Images

The steps for 3D images are almost similar with those for 2D images, except:

• use 3D patches as recommended for 3D images;

• there are three types of 3D output formats, “.mat”, “.nc” and “.tiff”, could be used;

• choose 3D as image dimension;