mnfpclassifier.m

用于目标识别的核辨别分析程序

% Modified Nearest Feature Plane Classifier-MNFP
function [MNFPCrate]=MNFPclassifier(features,test_features,trnum,tenum,classnum,K)

% features       the matrix that training samples projected on feature subspace(特征维数*训练样本数，列矢量)
% test_features  the matrix that test samples projected on feature subspace(特征维数*测试样本数，列矢量) 
% trnum          the number of training samples of each class
% tenum          the number of test samples of each class
% classnum       the number of classes
% K              the number nearest neighbors of each class

% MNFPCrate       the output correct classification rate of each class and the total
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% 首先计算距离矩阵：每一列为某测试点与所有训练像的距离
m=classnum*tenum; % 由于这里测试与训练像数相等，所以共用一个参数
dist=zeros(m,m);
for i=1:m
    for j=1:m
        dist(j,i)=norm(test_features(:,i)-features(:,j));
    end
end

for i=1:classnum
    % 对距离矩阵按降进行排序，并记住对应的训练像的位置
    [distnew1,ind1]=sort(dist((i-1)*trnum+1:i*trnum,:));
    ind(:,:,i)=ind1;
    featuresnew(:,:,i)=features(:,(i-1)*trnum+1:i*trnum);
    subdd(:,:,i)=test_features(:,(i-1)*tenum+1:i*tenum);
end

% number of NFP of each class
total=K*(K-1)*(K-2)/6;
% compute the parameter U (斜率) and the position P (垂点)
for m=1:classnum                 % 类别数
    for t=1:tenum                % 测试样本数/类
        for h=1:classnum
            for n=1:K
                subd(:,n,h)=featuresnew(:,ind(n,(m-1)*tenum+t,h),h);
            end
        end
        k=1;        
        for c=1:classnum             % 类别数
            for i=1:K                % 训练样本数/类
                for j=i+1:K          % 训练样本数/类
                    for h=j+1:K      % 训练样本数/类
                        Wijh=[subd(:,i,c),subd(:,j,c),subd(:,h,c)];
                        u=size(Wijh,2);
                        P=Wijh*inv(Wijh'*Wijh+(1e-3)*eye(u))*Wijh'*subdd(:,t,m);  % 垂足                        
                        dis(t,k,m)=norm(subdd(:,t,m)-P);            % 计算测试点到垂足的距离
                        k=k+1;       % 计算k条特征线(NFL)
                    end
                end
            end
        end       
    end    
end

% compute the correct recognition rate and the according mean value
CCrate=zeros(1,classnum);
for c=1:classnum
    dist=dis(:,:,c)';
    [Y,I]=min(dist);
    k=0;
    for i=1:tenum
        if ceil(I(i)/total)==c
            k=k+1;
        end
    end
    CCrate(c)=k/tenum;            
end
rtmean=sum(CCrate)/classnum;

% output
MNFPCrate=zeros(1,classnum+1);
MNFPCrate=[CCrate,rtmean];