Keras implementation of EEDS: End-to-End image super-resolution via deep and shallow convolutional networks

The original paper is end-to-end image super-resolution via deep and shallow convolutional networks

This implementation only focuse on two-scales' upsampling, and it is easy to change this code to perform upscaling with other ratio by change the strides in Conv2DTranspose. And because of the difference between the color space of YCrCb in Matlab and OpenCV, the final PSNR will have some difference. Results evluated with the MATLAB code will be a little higher.

Run the demo:

Excute: python EEDS_predict_demo.py

Train your own data:

Create your own data

open prepare_data.py and change the data path to your data

Excute: python prepare_data.py

training EEDS and test:

Excute: python EEDS.py

training EES and test:

Excute: python EES.py

Result:

EEDS and EES training for 200 epoches, with upscaling factor 2

Origin Image:

input:

 

EES:

 

EEDS:

 

feature map visualization of EES:

After conv1:

 

After Deconv:

 

This code of feature map visualization part refered the code of https://www.kaggle.com/abhmul/leaf-classification/keras-convnet-lb-0-0052-w-visualization