Calculating PSNR is a very important task in super-resolution research because PSNR is one of the most commonly used metrics. But early researchs were all on matlab, also including the code to calculate PSNR and preprocess(bicubic downsample and upsample). As most of the current mainstream deep learning frameworks are based of python, I need to re-write the calculating PSNR and preprocess code in python.

Thanks to the work in https://github.com/fatheral/matlab_imresize, it successfully reproduces the function imresize() in Matlab. This code is used to downsample and upsample the picture. And I found out that, the reproduction code is very successful when dealing with double type, but there are some problems when dealing with uint8 type(I only test the part of bicubic). The reason is still unknown yet.

My code for calculating PSNR is based on the code of matlab version in http://mmlab.ie.cuhk.edu.hk/projects/SRCNN.html.

Later I will draw a flowchart about calculating PSNR and put it on github.

Set5 has been prepared for test.

Briefly, you need to use the picture generate by the code in ./matlab_code and ./matlab_imresize-master, and put them into ./PSNR to calculate PSNR. You can read the readme.md in each code filter, which I write down how to use the code in detail.

Method 1 calculates PSNR without saving the pictures, while method 2 and 3 generate the pictures and then read the pictures to calculate PSNR.

Method 4 and 5 use the imresize() in https://github.com/fatheral/matlab_imresize.

Method 6 and 7 are an simple attempt to resize image with existing libraries, and get unsatisfied result.

I have tested the PSNR calculated by the pictures generated by Method 2 and 3 in both python and Matlab, and get the same result. So the code of calculating PSNR in python version is quite the same as the Matlab version.

Coming soon...