人脸验证算法Joint Bayesian详解及实现(Matlab)

python http://blog.csdn.net/cyh_24/article/details/49059475

github https://github.com/johnnyconstantine/Joint_Bayesian

实验所用的数据下载 http://www.down20.com/f-1036718061807849

下面为matlab代码，与论文的公式保持一致，运行需要比较大的内存。

addpath('../WDFdata')
%---------------------Load the training data------------------------
load('lbp_WDRef.mat')%there are  pictures,each contains  （*）
load('id_WDRef.mat')%persion id ,same id means same person
labels = id_WDRef;
clear id_WDRef;
X = double(lbp_WDRef);%trans int into double
clear lbp_WDRef;
train_mean = mean(X,);
X = bsxfun(@minus,X,train_mean);%substract the mean
% [COEFF,SCORE] = princomp(X,'econ');%PCA
[COEFF,SCORE] = pca(X,'econ');%PCA
train_x = SCORE(:,:);%
%----------------------------------------------------------
clear X;
clear SCORE;
%----------------------------------------------------------
%----------------------------------------------------------
%------------------------------load testing data-------------------------------------
load('lbp_lfw.mat')%it's 13233*5900
load('pairlist_lfw.mat')
%pairlist_lfw is an struct，it contains two * matrix，they are IntraPersonPair、ExtraPersonPair
normX = double(lbp_lfw);
train_mean = mean(normX,);
normX = bsxfun(@minus,normX,train_mean);
normX = normX * COEFF(:,:);%PCA
test_Intra = pairlist_lfw.IntraPersonPair;
test_Extra = pairlist_lfw.ExtraPersonPair;
%----------------------------------------------------------------------------------------------------------
clear lbp_lfw;
clear pairlist_lfw;
clear train_mean;
clear COEFF;
%-------------------------creat training pairs（intraPair和extraPair）---------------------------------------------------
[classes, bar, labels] = unique(labels);%labels
    nc = length(classes);%there are  person
    clear bar and classes;
train_Intra = zeros(nc*,);%*
for i=:nc%randperm（n，）
    train_Intra(*i-,:) = randperm(sum(labels == i),) + find(labels == i,,'first') - ;
    train_Intra(*i,:) = randperm(sum(labels == i),) + find(labels == i,,'first') - ;
end%choose two pictures from one person as intraPairs
train_Extra = reshape(randperm(length(labels),),,);%reshape *
train_Extra(labels(train_Extra(:,))==labels(train_Extra(:,)),:)=[];
train_Extra(size(train_Intra,)+:end,:)=[];
Dis_train_Intra = zeros(size(train_Intra,),);
Dis_train_Extra = zeros(size(train_Intra,),);
%------------------------------initialize before training---------------------------------------------------------
    m = length(labels); %
    n = size(train_x,);%
    [classes, bar, labels] = unique(labels);
    nc = length(classes);%为2995
    clear classes;
    clear bar;
    Sw = eye(size(train_x, ), size(train_x, ));
    Su = eye(size(train_x, ), size(train_x, ));%initialize sw su as identity matrix
    cur = {};%cell，*
    withinCount = ;
    numberBuff = zeros(,);% make sure that the largest number of a person is less than
    for i=:nc%
        cur{i} = train_x(labels == i,:);%make cell
        if size(cur{i},)>
            withinCount = withinCount + size(cur{i},);
        end
        if numberBuff(size(cur{i},)) ==
            numberBuff(size(cur{i},)) = ;
        end
    end%
    clear labels;
    u = zeros(n,nc);%*
    clear withinCount;
    oldSw = Sw;
    SuFG = cell(,);
    SwG = cell(,);%* cell
    %-------------------------------------EM-Like begin training-------------------------------------------------
    for k=: %set the cycle times
        F = inv(Sw);%formula（）in supplementary material
        ep =zeros(n,m);
        nowp = ;
        for g = :
            if numberBuff(g)==
                G = - .* (g .* Su + Sw) \ Su / Sw;%formula（）in supplementary material
                SuFG{g} = Su * (F + g.*G);%
                SwG{g} = Sw*G;
            end
        end%
        for i=:nc%
            nnc = size(cur{i}, );%number of each person
            u(:,i) = sum(SuFG{nnc} * cur{i}',2);%
            ep(:,nowp:nowp+ size(cur{i}, )-) = bsxfun(@plus,cur{i}',sum(SwG{nnc}*cur{i}',));%formula（）in supplementary material
            nowp = nowp+ nnc;
        end
        Su = cov(u');%su，
        Sw = cov(ep');%sw，。
        fprintf('%d %f\n',k,norm(Sw - oldSw)/norm(Sw));
        if norm(Sw - oldSw)/norm(Sw)<1e-%
            break;
        end
        oldSw = Sw;
    F = inv(Sw);
    mapping.G = - .* ( * Su + Sw) \ Su / Sw;%formula（）in supplementary material
    mapping.A = inv(Su + Sw) - (F + mapping.G);%formula（）
    mapping.Sw = Sw;
    mapping.Su = Su;
%---------------------------------begin testing-------------------------------------------------------------------

    for i=:size(train_Intra,)%formula（）
        Dis_train_Intra(i) = train_x(train_Intra(i,),:) * mapping.A * train_x(train_Intra(i,),:)' + train_x(train_Intra(i,2),:) * mapping.A * train_x(train_Intra(i,2),:)' -  * train_x(train_Intra(i,),:) * mapping.G * train_x(train_Intra(i,),:)';
        Dis_train_Extra(i) = train_x(train_Extra(i,),:) * mapping.A * train_x(train_Extra(i,),:)' + train_x(train_Extra(i,2),:) * mapping.A * train_x(train_Extra(i,2),:)' -  * train_x(train_Extra(i,),:) * mapping.G * train_x(train_Extra(i,),:)';
    end
    group_train = [ones(size(Dis_train_Intra,),);zeros(size(Dis_train_Extra,),)];%
    training = [Dis_train_Intra;Dis_train_Extra];%*
%-------------------------------------------------------------------------------------------------------------
    result_Intra = zeros(,);
    result_Extra = zeros(,);
    for i=:
        result_Intra(i) = normX(test_Intra(i,),:) * mapping.A * normX(test_Intra(i,),:)' + normX(test_Intra(i,2),:) * mapping.A * normX(test_Intra(i,2),:)' -  * normX(test_Intra(i,),:) * mapping.G * normX(test_Intra(i,),:)';
        result_Extra(i) = normX(test_Extra(i,),:) * mapping.A * normX(test_Extra(i,),:)' + normX(test_Extra(i,2),:) * mapping.A * normX(test_Extra(i,2),:)' -  * normX(test_Extra(i,),:) * mapping.G * normX(test_Extra(i,),:)';
    end%formula（）
    group_sample = [ones(,);zeros(,)];%
    sample = [result_Intra;result_Extra];%
%---------------------------------------
   % clear normX;%
%---------------------------------------
% classification
    [p,q]=size(group_train);%*
    starderd=;
    eg1=;
    num1=;
    eg2=;
    num2=;
    for i=:p
        if group_train(i,)==
            eg2=eg2+training(i,);%training
            num2=num2+;
        else
            eg1=eg1+training(i,);
            num1=num1+;
        end
    end
    starderd=(eg1+eg2)/(num1+num2);
    [p,q]=size(sample);%sample
    label=zeros(p,);%
    accuracy=;
    for i=:p
        if sample(i,)>starderd%same person
            label(i,)=;
        else
            label(i,)=;%not same person
        end
        if label(i,)==group_sample(i,)%
            accuracy=accuracy+;
        end
    end
    result = accuracy/p %
    if(result>=0.90)%if you can't reach this high accuracy you may end this by ctl + c or just set a lower number
        clear Sw and Su and F and G and numberBuff;%this should be done by the end of the for cycle
        break;
    end
end%new added by cx match the big for cycle
display('program done');