Robust Face Recognition via Sparse Representation

以前稀疏模型这块写过不少代码 都陆续找不到了 

Face Recognition小记

整理下代码

读此文J.Wright ... Yi Ma, “Robust Face Recognition via Sparse Representation”, PAMI.2009时

把PCA/LDA/Laplacian Face都写了一通，主要是降维上，上面三个在识别上都可以统一使用1-NN/linear-svm/near-subspace

相比：Robust Face Recognition via Sparse Representation(SRC)使用稀疏表示来做分类器，对降维不再敏感(子空间方法也不敏感)

后来发现个matlab降维工具包

我当时还把PCA/LDA/Laplacian写了一遍，费时。

以下仅贴SRC部分：

% test some thinking

clear;close all;clc
DatabasePath = 'C:\Users\CGGI003\Desktop\LiFeiteng\DataSets\CroppedYale';

newSize = floor([192 168]/5); %压缩Yale人脸 /5
[FaceData Label] = getFaceData(DatabasePath, newSize);

trainData = FaceData.train;
trainLabel = Label.train;
testData = double(FaceData.test);
testLabel = Label.test;
fprintf('...数据生成完成...\n')

fprintf('...SRC...\n')
% 使用降维工具包
no_dims = 200; %PCA降维
[trainData_R, mapping] = compute_mapping(trainData', 'PCA', no_dims);
trainData_R = trainData_R';
avg = mapping.mean';
testData_R = mapping.M'*testData;
avg = mapping.M'*avg;

% Robust Face Recognition via Sparse Representation(SRC)
SRC_Rate = Recognition_SRC(trainData_R, testData_R, trainLabel, testLabel,avg)

SRC部分 注意avg的使用

function  RecognitionRate = Recognition_SRC(A, testData, trainLabel, testLabel, avg)
%
%2013/05/21 绠楁硶缂栧啓娴嬭瘯
%2013/05/22 A -> B=[A I] + 娴嬭瘯

% Algorithm1 step2
A = double(A); testData = double(testData);
[m, n] = size(testData);
[~, N1] = size(A);
AA = A.*A;
norm_col = sqrt(sum(AA,1));
norm_col = repmat(norm_col,size(A,1),1);

A = A./norm_col;
%% 澧炲姞瀵归伄鎸＄殑澶勭悊
B = [A eye(m)];
K = max(size(unique(trainLabel)));
sigmaK = [];
for k = 1:K
    sigmaK(:,k) = ones(N1,1).*(trainLabel==k)';
end
estLabel = [];
h = waitbar(0,'Recognition...');

N = size(B,2);
for i = 1:n
    y = testData(:,i)-avg;
    
    w = SolveBP(B, y, N); %%SparseLab toolbox 
    residual = [];
    x = w(1:N1);
    for k =1:K
        %residual(k) = norm(y - A*(sigmaK(:,k).*x));  
        residual(k) = norm(y - A*(sigmaK(:,k).*x));         
    end
    [~,index] = min(residual);
    estLabel(1,i) = index;
    waitbar(i/n);
end
close(h);
RecognitionRate = sum(estLabel==testLabel)/max(size(testLabel));

end

PCA-200为时的正确率：

test 的系数矩阵 test数据是按类别依次排列的 可以看出 非零稀疏集中在对应类别上  跟预期效果符合

由于扩展了字典 [A I] 且使用的CroppedYale库 所以I部分对应的系数 很多非零 也可以理解了

字典A 可以做些处理 效果应该会更好些