While Downsampling 4:4:4 YCbCr image to 4:2:0 using imresize inbuilt function should we scale it by 0.5 or 0.25

I have converted RGB image into its YCbCr equivalent i.e. 4:4:4. And separated its Y, Cb and Cr components.

Now i need to downsample this 4:4:4 sample into 4:2:0 sample. For this i am using imresize inbuilt command.

Y component will be as it is in 4:2:0. The doubt i have is for Cb and Cr chroma components should i scale each component by 0.5 or 0.25.

Code:

%Load Image

RGB = imread('dbz.jpg');

%convert image to YCbCr color space

YCBCR = rgb2ycbcr(RGB);

%Extract Y and Cb, Cr components from the converted image it is in 4:4:4

Y=YCBCR(:,:,1);

Cb=YCBCR(:,:,2);

Cr=YCBCR(:,:,3);

Next which below code should i use 0.5 or 0.25 ?

%4:2:0 conversion of YCbCr using inbuilt command with bilinear

Cb_auto = imresize(Cb, 0.5, 'bilinear', 'Antialiasing', false);

Cr_auto = imresize(Cr, 0.5, 'bilinear', 'Antialiasing', false);

or

%4:2:0 conversion of YCbCr using inbuilt command with bilinear

Cb_auto = imresize(Cb, 0.25, 'bilinear', 'Antialiasing', false);

Cr_auto = imresize(Cr, 0.25, 'bilinear', 'Antialiasing', false);

Why i have this doubt is that i read like 4:2:2 samples chroma at 1/2th the luma samples.

Whereas 4:2:0 samples chroma at 1/4th the luma samples.

Help is appreciated!

Answer 1

[DGM]

Let me preface this answer by saying that I am not familiar with the conventions used for various file encodings. How exactly the image data gets downsampled (the method of interpolation, the use of antialiasing) is something I can't comment on.

That said, 4:2:2 reduces the amount of chroma information by half by reducing resolution by 0.5 in one dimension. Using 4:2:0 reduces the amount of chroma information by a quarter by reducing resolution by 0.5 in both dimensions. In other words, when you resize the chroma channels by 0.5, you're reducing the amount of pixels to 0.5^2 = 0.25.

[Sunil]

@DGM yes understood it now, by your explanation, Thank you!

So, i have to use this,

%4:2:0 conversion of YCbCr using inbuilt command with bilinear

Cb_auto = imresize(Cb, 0.5, 'bilinear', 'Antialiasing', false);

Cr_auto = imresize(Cr, 0.5, 'bilinear', 'Antialiasing', false);

The matrix i get after scaling Cb component by 0.5

Original image - 360x640 = 230,400 pixels

0.5 Scaled image - 180x320 = 57,600 pixels

230,400 * 1/4 = 57,600 (i.e. 1/4th of original pixels)

Thanks again!