www.pudn.com > image_mva_0.rar > dcKMeans.m

function [Classes,Centres,FinalDistance]=dcKMeans(Data,k,InitCentres,MaxIters) 
%DCKMEANS Performs k-means clustering 
% Classes=DCKMEANS(Data,k,InitCentres) where Data is a matrix with 
% variables in columns and records in rows, k is the number of  
% clusters to search for and InitCentres is a list of initial centres. 
% Classes=DCKMEANS(Data,k) randomly chooses initial centres from data set. 
% [Classes,Centres,FinalDistance]=DCKMEANS(Data,k,InitCentres) also returns 
% the centres and distances of each point to its nearest centre. 
 
% Copyright (C) David Corney 2000		D.Corney@cs.ucl.ac.uk 
 
if nargin < 3 
	InitCentres = ChooseInitialCentres(Data,k);		%randomly choose starting point (where needed) 
end 
Centres=InitCentres; 
OldCentres=Centres; 
if nargin<4 
   MaxIters=500; 
end 
 
[R,C]=size(Data); 
 
DataSq=repmat(sum(Data.^2,2),1,k);	%sum squared data - save re-calculating repeatedly later 
%Do we need DataSq? It's constant, and we're minimsing things... 
 
 
for i = 1:MaxIters 
    
   Dist = DataSq + repmat(sum((Centres.^2)',1),R,1) - 2.*(Data*(Centres'));   %i.e. d^2 = (x-c)^2 = x^2 + c^2 -2xc 
 
   [D,Centre]=min(Dist,[],2);		%label of nearest centre for each point 
 
   for j=1:k 
      idx=find(Centre==j); 
      if length(idx)>0 
         Centres(j,:)=mean(Data(idx,:)); 
      end 
   end 
    
   Change=sum(sum(abs(OldCentres-Centres))); 
   if Change < 1e-10	%Have we converged yet? 
      break 
   end 
   OldCentres=Centres; 
    
end 
 
 
[FinalDistance,Classes]=min(Dist,[],2);		%label points one last time