Compare images with svm

Hi,

I am trying to use SVM for comparison of two facial images based on extraction of various features. While SVM works fine for multi-class classification, comparing two images seems a bit more difficult. Basically, I have a set of pre-processed images of faces after detection and segmentation. The code is following:

%%Extract Features for training set 
training=imageSet('Y:\Bazy biometryczne\Face\att', 'recursive');
for i=1:3
        personA = 1;
        Pa=read(training(personA),i);
        Pa=Pa(:,:,1);
        % normalization
        Pa=double(Pa);
        Pa=(Pa-min(Pa(:)))/(max(Pa(:))-min(Pa(:)));
%         figure, imshow(Pa, [])
        trainingFeatures(i,:) = extractLBPFeatures(Pa);
        trainingLabel{i} = training(personA).Description;
end
for i=1:3
        personB = 2;
        Pb=read(training(personB),i);
        Pb=Pb(:,:,1);
        % normalization
        Pb=double(Pb);
        Pb=(Pb-min(Pb(:)))/(max(Pb(:))-min(Pb(:)));
%         figure, imshow(Pb, [])
        trainingFeatures(i+3,:) = extractLBPFeatures(Pb);
        trainingLabel{i+3} = training(personB).Description;
end        

%%train binary classification
% SVMModel = fitcsvm(X,Y,'KernelFunction','rbf','Standardize',true,'ClassNames',{'negClass','posClass'});
faceClassifier = fitcsvm(trainingFeatures,trainingLabel,'Standardize',true,'KernelFunction','linear' ); 

%%Test Images from Test Set 
personA1 = 1;
queryImage = read(training(personA1),4);
% normalization
queryImage=double(queryImage(:,:,1));
queryImage=(queryImage-min(queryImage(:)))/(max(queryImage(:))-min(queryImage(:)));
figure, imshow (queryImage,[])
queryFeatures = extractLBPFeatures(queryImage);
[personLabel, score] = predict(faceClassifier,queryFeatures)

As you can see, I am training the classifier with two sets (face of two persons) of 3 images. At the end I check if the test image is from one of the classes. Everything works great when I use an image from training. The problem appears when I use image of the trained person which was not used for training - it does not point correct person.

Generally, I tried with various different features but with no results. What would you suggest to me?

Answer 1

[kurakula sairam]

hey dude i have run this program and it works good after changing 1:3 to 1:4 for i value and can you tell me how to proceed this for N persons