基于MATLAB的人脸朝向识别

clear allclc;%% 人脸特征向量提取% 人数M=10;% 人脸朝向类别数N=5;% 特征向量提取pixel_value=feature_extraction(M,N);%% 训练集/测试集产生% 产生图像序号的随机序列rand_label=randperm(M*N);% 人脸朝向标号direction_label=repmat(1:N,1,M);% 训练集train_label=rand_label(1:30);P_train=pixel_value(train_label,:)';Tc_train=direction_label(train_label);T_train=ind2vec(Tc_train);% 测试集test_label=rand_label(31:end);P_test=pixel_value(test_label,:)';Tc_test=direction_label(test_label);%% K-fold交叉验证确定最佳神经元个数k_fold=10;Indices=crossvalind('Kfold',size(P_train,2),k_fold); error_min=10e10;best_number=1;best_input=[];best_output=[];best_train_set_index=[];best_validation_set_index=[];h=waitbar(0,'正在寻找最佳神经元个数.....');for i=1:k_fold% 验证集标号validation_set_index=(Indices==i);% 训练集标号train_set_index=~validation_set_index;% 验证集validation_set_input=P_train(:,validation_set_index);validation_set_output=T_train(:,validation_set_index);% 训练集train_set_input=P_train(:,train_set_index);train_set_output=T_train(:,train_set_index);for number=10:30for j=1:5rate{j}=length(find(Tc_train(:,train_set_index)==j))/length(find(train_set_index==1)) ;endnet=newlvq(minmax(train_set_input),number,cell2mat(rate));% 设置网络参数net.trainParam.epochs=100;net.trainParam.show=10;net.trainParam.lr=0.1;net.trainParam.goal=0.001;% 训练网络net=train(net,train_set_input,train_set_output);waitbar(((i-1)*21+number)/219,h);%% 仿真测试T_sim=sim(net,validation_set_input);Tc_sim=vec2ind(T_sim);error=length(find(Tc_sim~=Tc_train(:,validation_set_index)));if error<error_minerror_min=error;best_number=number;best_input=train_set_input;best_output=train_set_output;best_train_set_index=train_set_index;best_validation_set_index=validation_set_index;endendenddisp(['经过交叉验证，得到的最佳神经元个数为：' num2str(best_number)]); close(h);%% 创建LVQ网络for i=1:5rate{i}=length(find(Tc_train(:,best_train_set_index)==i))/length(find(best_train_set_i ndex==1));endnet=newlvq(minmax(best_input),best_number,cell2mat(rate),0.01);% 设置训练参数net.trainParam.epochs=100;net.trainParam.goal=0.001;net.trainParam.lr=0.1;%% 训练网络net=train(net,best_input,best_output);%% 人脸识别测试T_sim=sim(net,P_test);Tc_sim=vec2ind(T_sim);result=[Tc_test;Tc_sim]%% 结果显示% 训练集人脸标号strain_label=sort(train_label(best_train_set_index));htrain_label=ceil(strain_label/N);% 训练集人脸朝向标号dtrain_label=strain_label-floor(strain_label/N)*N;dtrain_label(dtrain_label==0)=N;% 显示训练集图像序号disp('训练集图像为：' );for i=1:length(find(best_train_set_index==1))str_train=[num2str(htrain_label(i)) '_'...num2str(dtrain_label(i)) ' '];fprintf('%s',str_train)if mod(i,5)==0fprintf('\n');endend% 验证集人脸标号svalidation_label=sort(train_label(best_validation_set_index)); hvalidation_label=ceil(svalidation_label/N);% 验证集人脸朝向标号dvalidation_label=svalidation_label-floor(svalidation_label/N)*N; dvalidation_label(dvalidation_label==0)=N;% 显示验证集图像序号fprintf('\n');disp('验证集图像为：' );for i=1:length(find(best_validation_set_index==1)) str_validation=[num2str(hvalidation_label(i)) '_'...num2str(dvalidation_label(i)) ' '];fprintf('%s',str_validation)if mod(i,5)==0fprintf('\n');endend% 测试集人脸标号stest_label=sort(test_label);htest_label=ceil(stest_label/N);% 测试集人脸朝向标号dtest_label=stest_label-floor(stest_label/N)*N;dtest_label(dtest_label==0)=N;% 显示测试集图像序号fprintf('\n');disp('测试集图像为：');for i=1:20str_test=[num2str(htest_label(i)) '_'...num2str(dtest_label(i)) ' '];fprintf('%s',str_test)if mod(i,5)==0fprintf('\n');endend% 显示识别出错图像error=Tc_sim-Tc_test;location={'左方' '左前方' '前方' '右前方' '右方'};for i=1:length(error)if error(i)~=0% 识别出错图像人脸标号herror_label=ceil(test_label(i)/N);% 识别出错图像人脸朝向标号derror_label=test_label(i)-floor(test_label(i)/N)*N;derror_label(derror_label==0)=N;% 图像原始朝向standard=location{Tc_test(i)};% 图像识别结果朝向identify=location{Tc_sim(i)};str_err=strcat(['图像' num2str(herror_label) '_'...num2str(derror_label) '识别出错.']);disp([str_err '(正确结果：朝向' standard...'；识别结果：朝向' identify ')']);endend% 显示识别率disp(['识别率为：' num2str(length(find(error==0))/20*100) '%']); % 特征提取子函数function pixel_value=feature_extraction(m,n)pixel_value=zeros(50,8);sample_number=0;for i=1:mfor j=1:nstr=strcat('Images\',num2str(i),'_',num2str(j),'.bmp');img= imread(str);[rows cols]= size(img);img_edge=edge(img,'Sobel');sub_rows=floor(rows/6);sub_cols=floor(cols/8);sample_number=sample_number+1;for subblock_i=1:8for ii=sub_rows+1:2*sub_rowsfor jj=(subblock_i-1)*sub_cols+1:subblock_i*sub_colspixel_value(sample_number,subblock_i)=...pixel_value(sample_number,subblock_i)+img_edge(ii,jj);endendendendendfunction [w1,w2]=lvq1_train(P,Tc,Num_Compet,pc,lr,maxiter)%% 初始化权系数矩阵% 输入层与竞争层之间权值bound=minmax(P);w1=repmat(mean(bound,2)',Num_Compet,1);% 竞争层与输出层之间权值Num_Output=length(pc);pc=pc(:);indices=[0;floor(cumsum(pc)*Num_Compet)];w2=zeros(Num_Output,Num_Compet);for i=1:Num_Outputw2(i,(indices(i)+1):indices(i+1)) = 1;end%% 迭代计算n=size(P,2);for k=1:maxiterfor i=1:nd=zeros(Num_Compet,1);for j=1:Num_Competd(j)=sqrt(sse(w1(j,:)'-P(:,i)));end[min_d,index]=min(d);n1=compet(-1*d);n2=purelin(w2*n1);if isequal(Tc(i),vec2ind(n2));w1(index,:)=w1(index,:)+lr*(P(:,i)'-w1(index,:));elsew1(index,:)=w1(index,:)-lr*(P(:,i)'-w1(index,:));endendendfunction [w1,w2]=lvq2_train(P,Tc,Num_Compet,lr,maxiter,w1,w2)%% 迭代计算n=size(P,2);for k=1:maxiterfor i=1:n% 计算各个竞争层神经元与当前输入向量的距离d=zeros(Num_Compet,1);for j=1:Num_Competd(j)=sqrt(sse(w1(j,:)'-P(:,i)));end% 寻找与当前输入向量距离最小的竞争层神经元标号，记为index1[min_d1,index1]=min(d);% 计算与index1相连接的输出神经元对应的类别a1_1=compet(-1*d);n2_1=purelin(w2*a1_1);a2_1=vec2ind(n2_1);% 寻找与当前输入向量距离次小的竞争层神经元标号，记为index2d(index1)=inf;[min_d2,index2]=min(d);% 计算与index2相连接的输出神经元对应的类别a1_2=compet(-1*d);n2_2=purelin(w2*a1_2);a2_2=vec2ind(n2_2);% 判断两个竞争层神经元对应的类别是否相等flag1=isequal(a2_1,a2_2);flag2=min_d1/min_d2>0.6;if ~flag1 && flag2if isequal(Tc(i),a2_1)w1(index1,:)=w1(index1,:)+lr*(P(:,i)'-w1(index1,:));w1(index2,:)=w1(index2,:)-lr*(P(:,i)'-w1(index2,:));elsew1(index1,:)=w1(index1,:)-lr*(P(:,i)'-w1(index1,:));w1(index2,:)=w1(index2,:)+lr*(P(:,i)'-w1(index2,:));endelsew1(index1,:)=w1(index1,:)+lr*(P(:,i)'-w1(index1,:));endendendfunction result=lvq_predict(P,Tc,Num_Compet,w1,w2)n=size(P,2);result=zeros(2,n);result(1,:)=Tc;for i=1:nd=zeros(Num_Compet,1);for j=1:Num_Competd(j)=sqrt(sse(w1(j,:)'-P(:,i)));endn1=compet(-1*d);n2=purelin(w2*n1);result(2,i)=vec2ind(n2);endNum_Correct=length(find(result(2,:)==Tc));accuracy=Num_Correct/n;disp(['accuracy=' num2str(accuracy*100) '%(' num2str(Num_Correct) '/' num2str(n) ')']);%% 清除环境变量clear allclc;%% 人脸特征向量提取% 人数M=10;% 人脸朝向类别数N=5;% 特征向量提取pixel_value=feature_extraction(M,N);%% 训练集/测试集产生% 产生图像序号的随机序列rand_label=randperm(M*N);% 人脸朝向标号direction_label=repmat(1:N,1,M);% 训练集train_label=rand_label(1:30);P_train=pixel_value(train_label,:)';Tc_train=direction_label(train_label);test_label=rand_label(31:end);P_test=pixel_value(test_label,:)';Tc_test=direction_label(test_label);%% 计算PCfor i=1:5rate{i}=length(find(Tc_train==i))/30;end%% LVQ1算法[w1,w2]=lvq1_train(P_train,Tc_train,20,cell2mat(rate),0.01,5); result_1=lvq_predict(P_test,Tc_test,20,w1,w2);%% LVQ2算法[w1,w2]=lvq2_train(P_train,Tc_train,20,0.01,5,w1,w2); result_2=lvq_predict(P_test,Tc_test,20,w1,w2);%% 清除环境变量clear allclc;%% 人脸特征向量提取% 人数M=10;% 人脸朝向类别数N=5;% 特征向量提取pixel_value=feature_extraction(M,N);%% 训练集/测试集产生% 产生图像序号的随机序列rand_label=randperm(M*N);% 人脸朝向标号direction_label=[1 0 0;1 1 0;0 1 0;0 1 1;0 0 1];train_label=rand_label(1:30);P_train=pixel_value(train_label,:)';dtrain_label=train_label-floor(train_label/N)*N;dtrain_label(dtrain_label==0)=N;T_train=direction_label(dtrain_label,:)';% 测试集test_label=rand_label(31:end);P_test=pixel_value(test_label,:)';dtest_label=test_label-floor(test_label/N)*N;dtest_label(dtest_label==0)=N;T_test=direction_label(dtest_label,:)'%% 创建BP网络net=newff(minmax(P_train),[10,3],{'tansig','purelin'},'trainlm'); % 设置训练参数net.trainParam.epochs=1000;net.trainParam.show=10;net.trainParam.goal=1e-3;net.trainParam.lr=0.1;%% 网络训练net=train(net,P_train,T_train);%% 仿真测试T_sim=sim(net,P_test);for i=1:3for j=1:20if T_sim(i,j)<0.5T_sim(i,j)=0;elseT_sim(i,j)=1;endendT_sim T_test。