How to randomly choose three pixels and use their values in kmeans

[KALYAN ACHARJYA]

Please refer the below link for your reference

https://in.mathworks.com/matlabcentral/answers/286264-k-means-for-a-grayscale-image#answer_677156

Answer 1

[Image Analyst]

See my attached example for how to do kmeans on a color image. In general kmeans is not a good color segmentation method. You'd probably be better off using discriminant analysis (like the other attached demo).

You might also be interested in the color frequency image code from the File Exchange:

https://www.mathworks.com/matlabcentral/fileexchange/28164-color-frequency-image?s_tid=srchtitle

[Image Analyst]

Did you look at the kmeans() documentation? Why are you using 'uniform' for the 'Start' option instead of the numerical array of the C points?

[Image Analyst]

I beleive there is still randomness in the process even though the starting points are the same. You won't necessarily get the same class indexes for all points all the time unless your clusters are well separated. What happened when you tried to pass in the values for those points like this

numberOfClusters = input('Enter the Number of clusters: ');
rgbImage = imread('peppers.png');
[rows, columns, numberOfColorChannels] = size(rgbImage)
imshow(rgbImage);
userPrompt = sprintf('Click %d points', numberOfClusters);
uiwait(helpdlg(userPrompt))
locations = ginput(numberOfClusters)
numPoints = size(locations, 1); % Should be 3 as long as you don't change the 3.
rgbAtLocations = zeros(numPoints, 3);
hold on;
for p = 1 : numPoints
    thisx = locations(p, 1);
    thisy = locations(p, 2);
    plot(thisx, thisy, 'r+', 'MarkerSize', 50, 'LineWidth', 2)
    rgbAtLocations(p, :) = impixel(rgbImage, thisx, thisy);
end
rgbAtLocations % Show in command window
% Make image into N by 3 matrix
[redChannel, greenChannel, blueChannel] = imsplit(rgbImage);
Z = [redChannel(:), greenChannel(:), blueChannel(:)];
% More code (if you want)..... Then:
[idx,C, SUMD ] = kmeans(double(Z), numberOfClusters, 'Start', rgbAtLocations);

[Image Analyst]

xout is the vector of class assignments. It's the class that it thinks your pixel should be in based on its color. So there is one element for however many pixels you have in your image. If you have a million pixels in the image, it is a million rows and one column. It is a 1-D array. It is not a 2-D array, nor should it be. And it certainly should not be a 3-D adday.

[Image Analyst]

If you want random points, then don't specify anything and I believe that's what it will use.

Xout is not the output of kmeans. It's the output of your function. So you need to take the class numbers and make RGB values. Like if the assigned class is 1 and you want 1 to be red, you'd have Xout(1, 1:3) = [1,0,0].

So you'd do

classColors = lines(numberOfClusters); % Whatever colormap you want.
assignedClasses =  kmeans(double(Z), numberOfClusters);
Xout = zeros(length(assignedClasses), 3); % RGB Image arranged in columns.
for k = 1 : length(assignedClasses) % for every pixel/color value.
    thisClass = assignedClasses(k); % Get this class
    Xout(k, :) = classColors(thisClass, :);
end
% Now turn from Nx3 matrix (R in col 1, G in col 2, B in col 3)
% into a 2D matrix -- a regular RGB image.
r = Xout(:, 1);
g = Xout(:, 2);
b = Xout(:, 3);
% Turn from column vectors into 2-D image
r = reshape(r, rows, []);
g = reshape(g, rows, []);
b = reshape(b, rows, []);
finalRGBImage = cat(3, r, g, b);

[Image Analyst]

kmeans starts its process with starting locations. If you don't specify them, I believe it picks them at random. So if you want random starting locations, you don't need to do anything. Of course you can specify starting locations if you have a good idea of where to start, but if you just pick random ones, it's just like the function picking random ones. So why pick random ones yourself when the function will do it for you if you don't specify the starting location?

The final image should be a classified image with 3 values: 1, 2, and 3, which you can map to any colors you want. It should not be a single scalar. Again, see my full demo for instructions.

Make sure you build Z like I did, not like you did.

Attach your image and m-file if you need more help.

[Image Analyst]

I think you forgot to attach face.jpg.