eigenfaces.m

Matlab 图像处理源码.(matlab源码)

% Author: A. I. Wilmer ([e-mail])
% September 2002 

TS_EXISTS=0;   % if 0 regenerate the training set and do SVD on it etc etc...

% establish the training set
directories = 35; trainingSamplesPerDirectory = 10;  % for directory-structured face database
trainingSamples = directories*trainingSamplesPerDirectory; fax = factor(trainingSamples);  plotHt = prod(fax,2)/fax(length(fax)); plotWd = fax(length(fax));

if (TS_EXISTS==0)
    disp(['Loading the training set (',num2str(trainingSamples),' images)...'])
    imageSize = 92;
    %exampleImage = double(imread('C:/Documents and Settings/aiw99r/My Documents/My Pictures/Brodatz Textures/D1.bmp','bmp'));
    trainingSet = zeros(imageSize,imageSize,trainingSamples);   
    m=1;
    for dirNumber = 1:directories
        for tspd=1:trainingSamplesPerDirectory
            
            exampleImage = double(imread(strcat('C:\Documents and Settings\aiw99r\My Documents\Face Databases\ATandT_Cambridge\s',num2str(dirNumber),'\',num2str(tspd),'.bmp'),'bmp'));
            
            %    exampleImage = double(imread(strcat('C:/Documents and Settings/aiw99r/My Documents/My Pictures/Faces/face',num2str(im(m)),'.bmp'),'bmp'));
            exIm = exampleImage(1:imageSize,1:imageSize); % crop to the desired training sample size
            trainingSet(:,:,m) = exIm;
            m=m+1;
        end
    end
    
    if(1==0)
        im_min = min(min(min(trainingSet))); im_max = max(max(max(trainingSet)));
        figure
        for m=1:trainingSamples
            subplot(plotHt,plotWd,m)
            imshow(trainingSet(:,:,m),[im_min im_max]),title(['Original Image ',num2str(m)])
        end
    end
    
    disp('Calculating mean image...')
    % generate mean image
    differenceImages = zeros(imageSize,imageSize,trainingSamples);
    meanImage = mean(trainingSet,3);
    figure,imagesc(meanImage),title('Mean Image'),colormap('gray')
    
    disp('Calculating difference images...')
    % subtract mean image to form the difference images
    for m=1:trainingSamples
        differenceImages(:,:,m) = trainingSet(:,:,m) - meanImage;
    end
    
    %figure
    %im_min = min(min(min(differenceImages))); im_max = max(max(max(differenceImages)));
    %for m=1:trainingSamples
    %    subplot(plotHt,plotWd,m)
    %    imshow(differenceImages(:,:,m),[im_min im_max]),title(['Difference Image ',num2str(m)])
    %end
    
    disp('Vectorising difference images')
    % form A -- which is a vector of the difference images
    A = zeros(imageSize^2,trainingSamples);
    for m=1:trainingSamples
        A(:,m) = reshape(differenceImages(:,:,m),imageSize^2,1);
    end
    
    % form A'A and find e-vectors and e-values
    %L = (A*A');
    %[eVecsOfL,eVals] = eig(L);    % V = full mx with columns being e-vectors, D = diagonal mx of e-values
    %eVecs = zeros(imageSize^2,trainingSamples);
    %eVecs = A*eVecsOfL;  % need to premultiply the e-vectors by A to get the e-vecs of A*A'
    
    disp(['Extracting eigen-values and eigen-images from SVD (of ',num2str(size(A,1)*size(A,2)),'-element matrix)...'])
    tic; [U,S,V] = svd(A,0); toc
    
    Eimage = zeros(imageSize,imageSize,trainingSamples);
    for i = 1:trainingSamples
        Eimage(:,:,i) = reshape(U(:,i),imageSize,imageSize);  % the columns of U are automatically the e-vectors of A*A' (i.e., the covariance matrix)
        Eval(i) = S(i,i)^2;   % S is a diagonal matrix of singular values, we need to square them to get eigen-values
    end
    if(1==0)
        figure
        im_min = min(min(min(Eimage))); im_max = max(max(max(Eimage)));
        for m=1:trainingSamples
            subplot(plotHt,plotWd,m)
            %    imshow(Eimage(:,:,m),[im_min im_max]),title(['Eigen-Image ',num2str(m),' with SVD'])
            imagesc(Eimage(:,:,m)),title(['Eigen-Image ',num2str(m),' with SVD']),colormap('gray')
        end
    end
    figure,plot(Eval),title('Eigen-values')
    
    % ---------------- END OF ASIDE ----------------------
    
    %eVecs = U(1:exImages,1:exImages);
    % perform linear combinations to get the eigen-images
    %eigenImage = zeros(imageSize,imageSize,trainingSamples);
    %for l=1:trainingSamples
    %    for k=1:trainingSamples
    %        eigenImage(:,:,l) = eVecs(k,l)*differenceImages(:,:,k);
    %    end
    %end
    %im_min = min(min(min(eigenImage))); im_max = max(max(max(eigenImage)));
    %figure
    %for l=1:trainingSamples
    %    subplot(plotHt,plotWd,l)
    %    imshow(eigenImage(:,:,l),[im_min im_max]),title(['E-Image ',num2str(l)])
    %   imagesc(eigenImage(:,:,l)),title(['E-Image ',num2str(l)]),colormap('gray')
    %   disp(['Eigenvalue ',num2str(l),' : ',num2str(Eval(l))]);
    %end
end
% ---------------------------------------- RECOGNITION TEST -------------------------------------------------

disp('Performing recognition...')
figure
% present a new face and attempt to recognise it
%recogniseImage = double(imread('C:/Documents and Settings/aiw99r/My Documents/My Pictures/Faces/face3.bmp','bmp'));
recogniseImage = double(imread(strcat('C:\Documents and Settings\aiw99r\My Documents\Face Databases\ATandT_Cambridge\s37\10.bmp'),'bmp'));
%recogniseImageTemp(1:121,1:160,1) = recogniseImage(1:121,2:161,1); recogniseImageTemp(1:121,161,1) = recogniseImage(1:121,1,1); recogniseImage=recogniseImageTemp; % implement a 1-pixel shift to investigate how well-posed the weights are

%recogniseImage = double(imread('C:/Documents and Settings/aiw99r/My Documents/My Pictures/Faces/lady1.bmp','bmp'));
recIm = recogniseImage(1:imageSize,1:imageSize); % crop to the desired training sample size
subplot(1,3,1),imagesc(recIm),title('Image to recognise')

% transform into eigenimage space
recDiffIm = recIm-meanImage;   % firstly, subtract the mean from it to get the difference image
subplot(1,3,2),imagesc(recDiffIm),title('Mean Difference Image'),colormap('gray')
weight = zeros(trainingSamples,1);
for m=1:trainingSamples
    weight(m) = (reshape(Eimage(:,:,m),imageSize^2,1))'*reshape(recDiffIm,imageSize^2,1);    % do inner product of each eigenimage with the difference image to get a weight vector
end
% use the weights to remake the image
reconstructionImage = zeros(imageSize,imageSize);
for m=1:trainingSamples
    reconstructionImage = reconstructionImage+(weight(m)*Eimage(:,:,m));    % should we incorporate in here (i.e., multiply weight(m) by eVals(m,m)????)
end
subplot(1,3,3),imagesc(reconstructionImage+meanImage),title('Eigen-image basis reconstruction'),colormap('gray')
figure,plot(weight),title('Weight vector for image to recognise')