Matlab 人脸识别之PCA算法,使用Yale人脸数据库
??? 写这个程序是老师布置的作业。一个莫名其妙的机会选了一个莫名其妙的课,于是写了与自己关系不大的人工智能的人脸识别的程序。这里给自己记录一下,估计这个学习都要和这个方面的打交道了。
???
??? Part 1:程序流程简介
???
??? 这个程序是典型的。在已有资源中使用一部分做训练集,找到一个合适的模型或者结论,然后用剩下的部分来测试自己的结论的正确度,进而一步步提高自己的算法效率或者正确性等。
?
??? 所以,在这次的程序中,前半部分是训练部分,中间有几段是画图部分,后面部分是测试部分。
?
??? Part 2:数据库和PCA算法简介
???
??? 数据库使用的是Yale的人脸数据库。一共15组图片,每组图片里面有11张图片。在我的程序里面,我使用了每组里面的8张照片为训练集,剩下的3张为测试集。所以,一共是120张训练照片,45张测试照片。
?
??? PCA算法步骤:
?
??? 1.Matrix X (input data)???????????????????????????????????????? N dimensional input space
???????即原始矩阵
?
??? 2.Matrix QX (Covariance of X)????????????????????????????? QX = cov(X) = E[(x-m)(x-m)T]
?????? 求出X的协方差矩阵QX
?
??? 3.Valuable λk(eigenvalue of QX)??????????????????????? λ1≥λ2≥λ3。。。
?????求出特征值,降序排列
?
?? 4.Vector Vk
???? 求出对应的特征向量
?
???5.Projection: V*X
???? 最后一步,在求出的向量基上面投影(降维)
?
??? Part 3:具体代码
???? ?%**********************************************% Aim : The first AI program homework% Title : PCA for face recognition% Author : GongWanlu & Hufei% Version : 1.0 final% Submit Time : 2011-04-07%**********************************************% %%%%%%%%%%%%%%%%%%%%%%INITIALclear allclcclose all% %%%%%%%%%%%%%%%%%%%%%%Some variables according to the Yale Face DBNum_subject = 15;Num_image = 11;Train_num_image = 8; %for every subject we choose 8 to trainTest_num_image = 9; %choose the left 3 images to test% %%%%%%%%%%%%%%%%%%%%%%Load DataData = [];for i=1:Num_subject for j=1:Train_num_image path = sprintf('FaceDB_yaleA/%03d/%02d.jpg',i,j); pic = imread (path); %read one picture %Make Data ,Add pic into Data pic_line = pic(1:147*137); %The pic size is 147*137 %pic_line is 1*N ,N=147*137. from up to %down,left to right. %Reshape 2D image to 1D image %vectors Data = [Data;double(pic_line)]; %add pic_line into Data endend% End of Load Data%%%%%%%%%%%%%%%%%%%%%%%Substract mean from Data and make covariance from centering Datasamplemean = mean(Data); %mean pic 1*Nfor k = 1:(Num_subject * Train_num_image) xmean(k,:)=Data(k,:)-samplemean; %Normalizeend %xmean is M*N ,each line is one pic %data(mean data) be normalizedsigma = xmean *xmean'; %M*M , here is 120*120[V D]=eig(sigma); %calculate the eigenvalue&eigenvector %eigenvalue in D,and vectors in V D1=diag(D); %the eigenvalues %%%%%%%%%%%%%%%%%%%%%%%% Sorting and eliminating eigenvalues % At first : sort desc Dsort=flipud(D1); Vsort=fliplr(V); %choose part eigenvalues Dsum = sum(Dsort); %sum of the eigenvalues,we only choose 80% %we have different ways to choose %eigenvalues we need,90%,or>1……temp_sum = 0;p = 0;while(temp_sum/Dsum<0.8) p = p+1; temp_sum = sum(Dsort(1:p));end%End of sort part %%%%%%%%%%%%%%%%%%%%%%%Train Step: get the coordinate system i=1; while(i<=p && Dsort(i)>0) face_base(:,i) = Dsort(i)^(-1/2) * xmean' * Vsort (:,i); i=i+1; end % Dsort(i)^(-1/2) used to normalize, make variance=1 % face_base is N*p % xmean' * Vsort (:,i); is change small matrix to big matrix. CHACHENG(Chinese) %next sentence is vary important is our train result allcoor = Data * face_base; %End of training %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%Draw part %draw CDF x = Dsort (1:p); x = flipud (x); for i=1:p y(i) = 1/p*i; end figure,plot(x,y,'r*'),hold on,plot(x,y,'b'); axis([0 inf 0 1]); title('CDF of selected values','fontsize',18); %draw the face (only use the first value) to show the result for k=1 temp = reshape(face_base(:,k),147,137); %1D to 2D end figure,imshow(mat2gray(temp)); %draw it title('one eigen face','fontsize',18);%draw the face (use 100 values) to show the result BigMap = []; for k=1:4 map = []; for j=1:4 temp = reshape(face_base(:,(k-1)*4+j),147,137);%1D to 2D map = cat(2,map,temp); end BigMap = cat (1,BigMap,map); end figure,imshow(mat2gray(BigMap)); %draw it Title ('Result of Chosen EigenValues','fontsize',18); %End of drwing part %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%Test Part Bingo = 0; for i=1:Num_subject for j=Test_num_image:Num_image %read the left 15*3 images path = sprintf('FaceDB_yaleA/%03d/%02d.jpg',i,j); test_pic = imread(path); test_pic_line = test_pic(1:147*137); test_pic_line = double(test_pic_line); tcoor= test_pic_line * face_base; %get the value for k=1:Num_subject* Train_num_image mdist(k)=norm(tcoor-allcoor(k,:)); %norm() distence end; %三阶近邻 [dist,index2] = sort(mdist); class1=floor( (index2(1)-1)/Train_num_image)+1; class2=floor((index2(2)-1)/Train_num_image)+1; class3=floor((index2(3)-1)/Train_num_image)+1; if class1~=class2 && class2~=class3 ID = class1; elseif (class1==class2) ID = class2; elseif (class2==class3) ID = class3; end; if ID==i Bingo=Bingo+1; end; end;end;accuracy = Bingo/45; %output the resultdisp(accuracy);% End of Test
?
附件里面,faceDB是使用的数据库。PCA是原始的程序。ProgramResult是一些简单的结论。直接把m文件和数据库文件夹放在一个目录下面就可以运行了。
?
?
?
?
?
1 楼 luxiaoling5566 2012-05-05 请问是怎么运行的呢是调用几张图片么
然后判断待识别图片与已知图片的相似性么 2 楼 Sunnie小食 2012-05-08 luxiaoling5566 写道请问是怎么运行的呢
是调用几张图片么
然后判断待识别图片与已知图片的相似性么
原则上确实是这样,不过不能这么说。这个是用特征去判断的,不是直接计算相似度的。
图片,不是调用,不能用这个词说 3 楼 luxiaoling5566 2012-05-11 Sunnie小食 写道luxiaoling5566 写道请问是怎么运行的呢
是调用几张图片么
然后判断待识别图片与已知图片的相似性么
原则上确实是这样,不过不能这么说。这个是用特征去判断的,不是直接计算相似度的。
图片,不是调用,不能用这个词说
但是,最后生成的三个图,第一个CDF图,代表什么意义呢,大二张图是第一个特征脸?第三个图那些脸是怎么选的呢,我看到有一句话temp = reshape(face_base(:,(k-1)*4+j),147,137);%1D to 2D,为什么那个k,j 要从1到4呢,最后得到了一个数是正确率么,有没有一个程序直接判断,未知图片属于跟已知图片的相似性的呢? 4 楼 luxiaoling5566 2012-05-20 请问最后结果是一张CDF,特征脸,识别率在哪儿可以看到呢