Can anyone help me to reduce number of for loops in this code ?

for n = 1:KMI
  for i = 1:M
      for j = 1:N
          for k=1:noC
           dis(k,1)=  sqrt(sum((reshape(im(i,j,:),[1,noF])-centers(k,:)) .^2,2));
          end 
       [minDis, index] = min(dis(1:noC,1));                
       im2d(i,j)=index;
       im_mind(i,j)=minDis;
       bestfit = sum(im_mind(:));
      end
  end
  plus=0;
  noValues=0;
  mean=zeros(noC,noF);
   for z=1:noF    
    for k=1:noC
      noValues=0;
      plus=0;
      for i = 1:M
          for j = 1:N
             if(im2d(i,j)==k)
                plus=plus+im(i,j,z);
                noValues=noValues+1;
             end
          end
      end
      if(noValues==0)
        mean(k,z)=0;
      else
        mean(k,z)=(plus/noValues);
      end
    end
   end
   if(centers==mean)
      break;
   end
   centers=mean;
 end

[Walter Roberson]

is im an array of coordinates, or is it an array of pixel values?

Is centers an array of rgb colors, or is an array of geometric coordinates?

Why are you not using rgb2ind() ?

[Akhil Movva]

im is an array of pixel values,

centers is an array of geometric coordinates,

actually i don't know this rgb2ind() funtion.

[Walter Roberson]

It does not make sense to compare color information im(i,j,:) to geometric locations.

It would make sense to compare color information im(i,j,:) to color information for a given list of colors, in order to determine the "closest" color. It happens that that is what rgb2ind() can do for you.

Answer 1

[Walter Roberson]

for n = 1:KMI
  for i = 1:M
      for j = 1:N
          for k=1:noC
           dis(k,1)=  sqrt(sum((reshape(im(i,j,:),[1,noF])-centers(k,:)) .^2,2));
          end 
       [minDis, index] = min(dis(1:noC,1));                
       im2d(i,j)=index;
       im_mind(i,j)=minDis;
       bestfit = sum(im_mind(:));
      end
  end

You overwrite bestfit in every iteration of j inside every iteration of i inside every iteration of n . If you look at the rest of your code, you will find that you never use bestfit afterwards. This means that you only want the final value to use afterwards for some reason -- the value associated with n = KMI, i = M, j = N. So you can save resources by not doing the calculations for bestfit inside the loops, and instead do it afterwards with n=KMI, i=M, j=N only.

You are overwriting im_mind(i,j) inside for n for i for j, and you do not use im_mind inside the for n loop, except to calculate bestfit, which as discussed above, you do not use inside any of the loops. You can save resources by not doing the calculation inside the for n loop. Instead, just calculate im_mind once, with n=KMI, and the entire range of i and j values, creating only one single im_mind array. Then after that has been done, you can calculate bestfit once as discussed above. More likely you do not need im_mind or bestfit at all.

You do do further computation with the im2d index array. As I described in the comments, you can build it by a single call to rgb2ind()

I do not immediately follow the code for the loops after that point. It looks like for each color, you are calculating the mean R, G, or B component from the image from the pixels that were considered to match that color. If that is correct, then there much more efficient ways to do that:

  for K = NoC : -1 : 1
     output_mean(:,K) = accumarray(im2d(:), reshape(im(:,:,K), [], 1), [NoC,1], @mean, 0);
  end
  

Then you have

     if(centers==mean)
      break;
   end
   centers=mean;
   

Is it possible that all of this code is an attempt to find the noC "best" colors to partition the image into? If so, then you can replace the entire code with a call to rgb2ind(im, noC)

[Akhil Movva]

Thankyou for you answering. It was very helpful. The above code is part of k-means clustering program for 3 feature and 5 feature image

[Walter Roberson]

Just call kmeans() directly instead of programming it yourself.