knn.m

传感器网络探测声信号通过特征提取进行目标设别分类

function [class]=knn(Pr,Tr,Pt,kN)
% Usage: [Cmat,C_rate]=knn(Pr,Tr,Pt,Tt,kN)
% kNN- k-Nearest Neighbor Classifier
% copyright 1993-1996 by Yu Hen Hu
% Last revision 2/12/03 by Marco F. Duarte
% Pr:  training feature vector  K x N (prototype), the classifier
% Tr:  training target vector   K x S (labels of prototye)
% Pt:  testing/validating feature vector Q x N 
% kN:   # of nearest neighbors used.
% Assume L2 norm distance
% Cmat:  confusion matrix of the testing/validation set
%      = classified class(Sx1) * true classes (1xS)
% C_rate: classification rate (%)
% This is done so that attenuation and distance have approx. same weight.
Pr(:,2) = Pr(:,2)/500;
Pt(:,2) = Pt(:,2)/500;

[K,S]=size(Tr);
[Q,N]=size(Pt);
rbias=sum(Pr'.*Pr')*.5;  %  1/2||W||^2    1 by K
if S==1,
  oneS=eye(S+1);
  Tr=[Tr ones(K,1)-Tr];  %  if S=1, change Tr,Tt to 2 class format
else
  oneS=eye(S);
end

% calculate L2 norm distance
class=[];
for i =1:Q,
   d =-Pr*Pt(i,:)'+rbias';   % d is K by 1, distance
   %  choose the minimum kN terms (k-nearest neighbor) of each column
   [y,idx]=sort(d);   % both y and idx are K by 1, first kN entries of idx
                      % gives the indices of the kN nearest neighbor
   if kN > 1,
      [yy,kidx]=max(sum(Tr(idx(1:kN),:)));
      class=[class;oneS(kidx,:)];
   else
      class=[class;Tr(idx(1:kN),:)];
   end
end % i-loop
if S == 1,
    class = class(:,1);
end