Computing pixels value(Uk,Uj) from U -Image in YUV color Space

Hi everyone

I want to compute the pixel values let say Uk and Uj from UImage.Here is my code in which I am doing indexing.

UImage=imread(Uimage);

B=ones(length(UImage));

slidingwindow=conv2(B,UIimage,'same');%this part computes only centre pixels.

[pixelrow pixelcol]=size(slidingwindow);%Pixelrow and pixelcol are center pixels of window

for windrow=-1:1

row=pixelsrow+windrow;

for windcol=-1:1

col=pixelscol+windcol;

%% here I need some piece of code further May be Uk=UImage(row,col)???????????

end end

[iup geii amiens]

Hello Sir,

I am working on the same sheet. Please can i have your code and we may help each other :D

I will be grateful Have a nice day

Answer 1

[Thorsten]

To get the pixel values of a gray scale image

 I = imread('cameraman.tif');
 x = 20; y = 30; 
 Ixy = I(y, x);

To get the G pixel value of an RGB image

 I = imread('peppers.png');
 IGxy = I(y, x, 2);

Same should work for your U image if you have stored just the U channel or all three channels in your image.

[Algorithms Analyst]

My Dear this is very basic concept.Pleas read my question properly as I need something else.I have already got U Image now I need to compute the pixel values from different locations thats why I applied sliding window concept..Please read it properly thanku.

[Thorsten]

If you want to compute some function, e.g., mean on sliding blocks, you can use

 M = colfilt(I, [3 3], 'sliding', @mean);

If this does not help you, please ask more precisely what you want to compute.

[Algorithms Analyst]

Ok for your understanding I explain it more clearly..

I want to compute Gussian kernel of UV image in YUV space.I have separated out UV image form YUV space.Now I want to compute Gaussain Kernel by using the formulae below.

DU=Uk-j and DV=Vk-Vj

Gaussian=1./(2*sigma*sigma)*exp(-Du*Du+Dv*Dv./(2*sigma));

Thats why I need to compute the pixel values of U image from j and k location.therefore I used sliding window concept here but I am stuck on this point.

Thanks

[Thorsten]

First you get the n x n subimages Uk and Vk from U and V at position x, y

   Uk = U(ind1, ind2);
   Vk = V(ind1, ind2);

with

   ind1 = y-n/2:y+n/2;
   ind2 = x-n/2:x+n/2;

The contribution of these subimages to bin j asscociated with some fixes colors Uj, Vj is then

   dU = Uk(:) - Uj;
   dV = Vk(:) - Vj;

   Gj = Kj*exp(-(Du.^2+Dv.^2)/(2*sigma^2));

with normalization coefficient Kj.

[Algorithms Analyst]

can you please tell me what is x,y and n and how to compute them?

I have two images U and V.can you please tell me how can we compute ind1 and ind2? thank you.

[Algorithms Analyst]

As I applied the sliding window concept in my code refered by image analyst.But I did not see any of sliding window in your piece of code.Can you plz explain little bit more.

Thank you.

[Algorithms Analyst]

and what is Uj and Vj and how did you compute it?

[Thorsten]

I assume that you refer to the paper http://philippe.noriega.free.fr/fichiers/visapp06.pdf

Then x and y are the location in the image where you compute your histograms. n = 12 is the size of your subwindow that is applied with a gap of 3 to your image.

To slide your window, use

gap = 3;
for x = 1:gap:size(I, 1)

    for y = 1:gap:size(I, 2)

      ind1 = y-n/2:y+n/2;
      ind2 = x-n/2:x+n/2;

And for each location you need a third loop over the number of bins in your histogram

      for j = 1:Nbins

        dU = Uk(:) - Uval(j);
        dV = Vk(:) - Vval(j);

        Gj = Kj*exp(-(Du.^2+Dv.^2)/(2*sigma^2));

        CH(x, y, j) = Gj;
      end
    end
  end

Answer 2

[Image Analyst]

Your code is all wrong. Try this:

UImage=imread(Uimage);
windowSize = 5;
sigma = whatever.....
kernel = fspecial('gaussian', windowSize, sigma);
slidingwindow=conv2(UIimage, kernel ,'same');

You don't need double for loops after that - conv2 does the sliding window filtering so it's already done. No need to try to do it again manually.

[Algorithms Analyst]

Yes.this work fine but as you told in your previous answer here (<http://www.mathworks.co.kr/matlabcentral/answers/63950-computing-color-gaussian-kernel)>.

we need to compute Uk and Uj pixel from different locations thats why I was working on it.So is this code the Color Gassian kernel is (slidingwindow=conv2(UImage,kernel,'same')) ?So where is Uj and Uk?

As slidingwindow contains the centre part...

Thanks ....

[Thorsten]

Based on one of AlgorithmAnalyst's previous question I think that his question refers to the local kernel color histogram algorithm described in http://philippe.noriega.free.fr/fichiers/visapp06.pdf

[Algorithms Analyst]

Yes you are right.This is what I am implementing.But confused too much As Image Analyst code I cannot see Uk and Uj.May if I am not wring these two pixels location are in sliding window.?

[Thorsten]

As far as I have understood the algorithm, you compute histograms at different image locations x, y. The histograms are by counting the frequency of Nj colors (Uj, Vj) in a 12 x 12 window around (x, y). To make the counting more robust, you increment the (Uj, Vj) bin not only if you encounter the color (Uj, Vj) in your 12 x 12 window, but you compute a Gaussian weighted similary between your 12 x 12 patch and the color (Uj, Vj). See my code I sketched above.

[Algorithms Analyst]

Yes and I have commented on your question how to compute these locations xy and n and how to compute Uj and Vj this is my question.But I think Image Alalyst has solved it by directly calling fspecial and conv2 finctions in Matlab.

Thanks for your concern..

[Image Analyst]

The conv2 code essentially is 4 nexted for loops, the outer two move the window along - they scan the image over all rows and all columns. At a particular row and column inside those two loops there is a nested pair that scans over the pixels in the window (kernel) that has your Gaussian weights. Let's say these loops are over rw and cw. For example, if you have a 3 by 3 window, the outer loops will (eventually) get you to row,column 135, 154. So we need to scan a window around there from row 134 to 136, and from column 153 to 155. So when it's doing that, j and k are calculated from row and rw and column and cw. Specifically if rw goes from -1 to +1, then j = row + rw, and k = column + cw. So j and k would start out at (134, 153), then go to (135,153) then to (136,153) then to (134, 154) and so on until it visits all 9 locations in the window centered around (135, 155). But when you're taking the U value from (134,154), you don't have a Gaussian with those values because the Gaussian is just a 3x3 window so the value of the Gaussian at that upper left window pixel would be the Gaussian value at (-1, -1), in other words, the Gaussian value at a distance of sqrt(2) away from the center of the Gaussian. I hope that explains it better.

[Algorithms Analyst]

So slidingwindow is the required color gaussian kernel?i dont need further calculation?

[Image Analyst]

Correct, but I haven't really tried to understand what the algorithm does. You're just blurring the U channel. Not sure if that's part of the algorithm or not - I'm just going by what you said.

[Algorithms Analyst]

Yeah. That's so interesting.In the color Gaussian kernel part of the algorithm two other pixels values from i and j locations are computed and they are VJ and Vk respectively.So I think they are taking just UV image from YUV colorspace by ignoring Y channel and computing the color gaussain kernel in UV color space.But anyway thanks for your time and concern.

Thanksssss......cheers..