Implement VDSR with TensorFlow

Introduction

We present a highly accurate single-image super-resolution (SR) method, learn residuals only and use extremely high learning rates enabled by adjustable gradient clipping

Environment

• Ubuntu 16.04
• Python 2.7.12

Dependency

pip

• Tensorflow
• Opencv
• h5py

If you meet the problem with opencv when run the program

libSM.so.6: cannot open shared object file: No such file or directory

please install dependency package

sudo apt-get install libsm6
sudo apt-get install libxrender1

How to train

python main.py

if you want to see the flag

python main.py -h

trainning with 10 layer (default is 20)

python main.py --layer 10

How to test

If you don't input a Test image, it will be default image

python main.py --is_train False

then result will put in the result directory

if you want to see the result with custom layer ,ex: 10 layer

python main.py --is_train False --layer 10

If you want to Test your own iamge

use test_img flag

python main.py --is_train False --test_img Train/t20.bmp

then result image also put in the result directory

Result

• Origin

  

  

• Bicbuic

  

  

• Result

  

  

Reference

kweisamx/SRCNN