pca.rar_PCA 图像_matlab 图像特征提取_pca_特征提取

%% By hhh 2008-3-6 %% TITLE: PCA for the preprocessed images %% Read the training data clear all close all set(0,'RecursionLimit',100000); % training = dir('..\..\..\data\aurora_data\arc_training'); % path = '..\..\..\data\aurora_data\arc_training'; % I try the program using face data first % for MIT database % path = 'E:\FaceDetection\MyData\MIT\Training'; % for Viola database----only here has difference! and LoadI.m one place different from MIT. path = 'E:\FaceDetection\MyData\Viola\Training'; % Load images [A,trainFNum,trainNonFNum,row,col] = LoadImgs(path); trainingNum = trainFNum + trainNonFNum; % obtain the diffirent images between the images and the average image [A,FAImean] = CalDiffirentImg(A,trainingNum,row,col); %% PCA for the training images % (1) covariance matrix, S = A*A' (skip computing by using svd) % (2) partial eigenvalue decomposition S = U'*E*V % [U,E,V] = svd(A); % singular val decomp much faster------Out of memory!!! CovOfDiffImg = A*A'; % according to Turk,calculate A'*A firstly % in the condition of m<<n^2(m is faces number,n^2 is their size), then eigenVecs=A*U. % But here, the images are only 19*19,then A is 361*1000,so m>n^2. % So,directly A*A'.------hhh, 3.31,2008 [U,E,V] = svd(CovOfDiffImg); % singular val decomp much faster eigVals = diag(E); TotalEnergy = sum(eigVals)*0.98; for i = size(eigVals,1):-1:1 if(sum(eigVals(1:i))<=TotalEnergy) pcaNum = i + 1; break; end end % (3) get sorted eigenvalues (diag of E) and eigenvectors (U) % try diag(E.^2) later!!!!!! eigVecs = U(:,1:pcaNum); % these are eigenfaces --- not A*U'!! ----hhh % show the eigenfaces % titleName = 'EigenAuroras'; % titleName = 'EigenFaces'; % ShowSubplot(titleName,eigVecs,row,col); %% Projection of test images onto base space % the projected coefficients of the training set A.every column is 'pcaNum' % coefficients to each image in the training set(each column of A). WeightsOfTrainingSet = eigVecs'*A; % see a example: figure;imshow(reshape(uint8(eigVecs*WeightsOfTrainingSet(:,66)),row,col)); figure;imshow(reshape(uint8(eigVecs*WeightsOfTrainingSet(:,66)+FAImean),row,col)); a = uint8(eigVecs*WeightsOfTrainingSet(:,66)+FAImean); W = U'*A; figure;imshow(reshape(uint8(U*W(:,66)+FAImean),row,col)); aa = uint8(U*W(:,66)+FAImean); %% Read the test data % test = dir('..\..\..\data\aurora_data\arc_test'); % path = '..\..\..\data\aurora_data\arc_test'; % I try the program using face data first % for MIT database % path = 'E:\FaceDetection\MyData\MIT\Test'; % for Viola database path = 'E:\FaceDetection\MyData\Viola\Test'; % Load images [B,testFNum,testNonFNum,row,col] = LoadImgs(path); % obtain the diffirent images between the test images and the previous % average image of the training set A. for j = 1:size(B,2) B(:,j) = (B(:,j) - FAImean); %figure;imshow(reshape(uint8(A(:,j)),m,n)); end %% Projection of test images onto base space % the projected coefficients of the test set B.every column is 'pcaNum' % coefficients to each image in the test set(each column of B). WeightsOfTestSet = eigVecs'*B; % see a example: figure;imshow(reshape(uint8(eigVecs*WeightsOfTestSet(:,66)),row,col)); figure;imshow(reshape(uint8(eigVecs*WeightsOfTestSet(:,66)+FAImean),row,col)); b = uint8(eigVecs*WeightsOfTestSet(:,66)+FAImean); WB = U'*B; figure;imshow(reshape(uint8(U*WB(:,66)+FAImean),row,col)); bb = uint8(U*WB(:,66)+FAImean); %% Calculate the distances between the training set itself and % the distances between the training set and test set. %belong to different database % 1~3 database % TrainLableMatrix = [zeros(1,500),ones(1,500)]; % TestLableMatrix = [zeros(1,500),ones(1,500)]; % 4\5\6 database % TrainLableMatrix = [zeros(1,1000),ones(1,1000)]; % TestLableMatrix = [zeros(1,1000),ones(1,1000)]; % 7 database % TrainLableMatrix = [zeros(1,1500),ones(1,1000)]; % TestLableMatrix = [zeros(1,1500),ones(1,1000)]; % 8 database % TrainLableMatrix = [zeros(1,1500),ones(1,1200)]; % TestLableMatrix = [zeros(1,1500),ones(1,1200)]; % general database!!! TrainLableMatrix = [ones(1,trainFNum),zeros(1,trainNonFNum)]; TestLableMatrix = [ones(1,testFNum),zeros(1,testNonFNum)]; % Mr. Vassilis Athitsos said in his code readme.txt: It is strongly % recommended that you randomize the order of your training objects before % you creat these matrice.----TestTrainMatrix,TestLableMatrix aren't changed. RandTrainLableMatrix = zeros(size(TrainLableMatrix)); RandWeightsOfTrainingSet = zeros(size(WeightsOfTrainingSet)); % generate 'trainingNum' random numbers index = []; CountNum=0; % The following 7 lines is good for selecting 'trainingNum' random numbers too,but I use an other more simple way!!! % while (trainingNum-CountNum)>500 % [index,c1] = GenRandArr(index,trainingNum-CountNum,trainingNum); % CountNum = CountNum + c1; % end % for i = CountNum+1:trainingNum % index(i) = GenRandNum(index,i,trainingNum); % end n = 3; while trainingNum>CountNum n = n + 1; [index,CountNum] = GenRandArr(index,trainingNum*n,trainingNum); end for i = 1:trainingNum RandTrainLableMatrix(i) = TrainLableMatrix(index(i)); RandWeightsOfTrainingSet(:,i) = WeightsOfTrainingSet(:,index(i)); end TrainTrainMatrix = dists(RandWeightsOfTrainingSet,RandWeightsOfTrainingSet); TestTrainMatrix = dists(WeightsOfTestSet,RandWeightsOfTrainingSet); %% Write the two kinds of matrixs into traintrain_distances.bin and traintest_distances.bin filename1 = 'E:\GaitManifoldLearning\Boostmap\boostmap_code\boostmap\data\bm_datasets\FaceData\traintrain_distances.bin'; result1 = write_float_matrix(TrainTrainMatrix, filename1); filename2 = 'E:\GaitManifoldLearning\Boostmap\boostmap_code\boostmap\data\bm_datasets\FaceData\testtrain_distances.bin'; result2 = write_float_matrix(TestTrainMatrix, filename2); % Write the train and test lables matrixs into training_lables.bin and % test_labless.bin ----- 0 for nonface while 1 for face filename3 = 'E:\GaitManifoldLearning\Boostmap\boostmap_code\boostmap\data\bm_datasets\FaceData\training_labels.bin'; result = write_float_matrix(RandTrainLableMatrix, filename3); filename4 = 'E:\GaitManifoldLearning\Boostmap\boostmap_code\boostmap\data\bm_datasets\FaceData\test_labels.bin'; result = write_float_matrix(TestLableMatrix, filename4); %% Turk's face recognition % % pesi_database_mediati is a matrix with the averaged eigenface components of the images % % present in database. Each class has its averaged eigenface. % pesi_database=zeros(pcaNum, dbInfor.individualNum); % numero_elementi_classe=zeros(dbInfor.individualNum,1); % for ii=1:dbInfor.faceNum % ingresso_database=double(dbInfor.data{ii,1}); % classe_database=dbInfor.data{ii,2}; % pesi_correnti=eigFaces'*(ingresso_database-media); % pesi_database(:,classe_database)=pesi_database(:,classe_database)+pesi_correnti; % numero_elementi_classe(classe_database)=numero_elementi_classe(classe_database)+1; % end % for ii=1:dbInfor.individualNum % pesi_database_mediati(:,ii)=pesi_database(:,ii)/numero_elementi_classe(ii); % end % % weight.individualNum = dbInfor.individualNum; % weight.pcaNum = pcaNum; % weight.Vtrue = eigFaces; % weight.Dtrue = eigVals; % weight.media = media; % weight.meanFaces = pesi_database_mediati; % weight.individualName= dbInfor.individualName;