How to correctly calculate PSNR of yuv420p video?

Hi, I am working on calculation of psnr for yuv420p formatted video. I found a sample code here: http://www.mathworks.com/matlabcentral/fileexchange/12455-psnr-of-yuv-videos . The code works fine for mse. However, I have one question on calculation of PSNR from MSE. If the MSE is calculated for individual component y, u or v which are individually 8 bits per sample, then the psnr is correctly calculated as 10*log10((2^8 - 1)/mse). But when taking whole 'yuv' for the mse, is it a total mse or average mse for y, u and v components? If we consider the downsampling of chroma components, say for yuv420p format, total or average bits per pixel will not be 8 (generally 12 or 16 bits total). Can anyone who has knowledge of this yuv format please clarify how to calculate psnr for yuv420p? Thanks in advance.

[Kejun Wu]

replace

       %% compute the psnr 
%         if msemean ~= 0
%             PSNR = 10*log10((255^2)/msemean)
%         else
%             PSNR = Inf;
%         end

to

       for cntf = 1:framenumber1
           if msemean ~= 0
                PSNRframe(cntf) = 10*log10((255^2)/mse(cntf));
           else
                PSNR = Inf;
           end
       end
       PSNR = (sum(PSNRframe)/length(PSNRframe));

[Image Analyst]

Since cntf is an integer (it's the loop iterator variable),

10*log10((255^2)/mse(cntf));

will have nothing at all to do with the current frame's contents. And since msemean doesn't depend on cntf, it will always take the same path through the "if" for any iteration.

[Karbala'a Unvi. Science]

Hi Dear Image Analyst

Is there any short way to calculate the PSNR and MSE for the video, than this one.

If you have an video to explain that I am so happy to see it

I am waitting for an answer from you.

Thank you in advance

Zee

[Image Analyst]

There are now functions in the Image Processing Toolbox called immse() and psnr(). You could use those.

Answer 1

[Image Analyst]

When you pass in 'yuv', it does YUV1(:)-YUV2(:) so it does it element by element over every pixel in the image and averages them all together. It doesn't make a lot of sense since the scale and range of the Y axis is different than that of the U and V axes. You'd be better off calculating the CIE DeltaE 2000, which is more relevant to the way people see.

By the way, you don't need a File Exchange program since there is immse() and psnr() built in to the Image Processing Toolbox now.