opencv学习之路（40）、人脸识别算法——EigenFace、FisherFace、LBPH

一、人脸识别算法之特征脸方法（Eigenface）

1、原理介绍及数据收集

特征脸方法主要是基于PCA降维实现。

详细介绍和主要思想可以参考

http://blog.csdn.net/u010006643/article/details/46417127

上述博客的人脸数据库打不开了，大家可以去下面这个博客下载ORL人脸数据库

http://blog.csdn.net/xdzzju/article/details/50445160

下载后，ORL人脸数据库有40个人，每人10张照片。

2、流程

3、相关图示

 

4、代码

#include <opencv2/opencv.hpp>
#include <opencv2/face.hpp>

using namespace cv;
using namespace cv::face;
using namespace std;

//对原图归一化
Mat normal(Mat src, Mat dst) {
    if (src.channels() == 1)//若原图单通道
        normalize(src, dst, 0, 255, NORM_MINMAX, CV_8UC1);
    else //否则，原图三通道
        normalize(src, dst, 0, 255, NORM_MINMAX, CV_8UC3);
    return dst;
}

void main() {
    //读取文件，转换为数据流
    string filename = string("at.txt");
    ifstream file(filename.c_str(), ifstream::in);
    if (!file)
        cout << "error" << endl;

    string line, path, classlabel;
    vector<Mat>image;
    vector<int>labels;
    char separator = ';';
    while (getline(file,line))
    {
        stringstream liness(line);
        getline(liness, path, separator);
        getline(liness, classlabel);
        if (!path.empty()&&!classlabel.empty())
        {
            //cout << "path:" << path<< endl;
            image.push_back(imread(path, 0));
            labels.push_back(atoi(classlabel.c_str()));
        }
    }

    if (image.size() < 1 || labels.size() < 1)
        cout << "invalid image path..." << endl;

    int height = image[0].rows;
    int width = image[0].cols;
    //cout << "height:" << height << "," << width<<endl;

    //最后一个人为测试样本
    Mat testSample = image[image.size() - 1];
    int testLabel = labels[labels.size() - 1];
    image.pop_back();
    labels.pop_back();

    //训练
    Ptr<BasicFaceRecognizer>model = createEigenFaceRecognizer();
    model->train(image, labels);

    //识别
    int predictLabel = model->predict(testSample);
    cout << "actual label:" << testLabel << ",predict label:" << predictLabel << endl;

    //获得特征值，特征向量，均值    平均脸
    Mat eigenvalues = model->getEigenValues();
    Mat eigenvectors = model->getEigenVectors();
    Mat mean = model->getMean();
    Mat meanFace = mean.reshape(1,height);
    Mat dst;
    dst= normal(meanFace,dst);
    imshow("Mean Face", dst);
    
    //特征脸
    for (int i = 0; i < min(10,eigenvectors.cols); i++)
    {
        Mat ev = eigenvectors.col(i).clone();
        Mat eigenFace = ev.reshape(1, height);
        Mat grayscale;
        grayscale = normal(eigenFace, grayscale);
        Mat colorface;
        applyColorMap(grayscale, colorface, COLORMAP_BONE);
        char* winTitle = new char[128];
        sprintf(winTitle, "eigenface_%d", i);
        imshow(winTitle, colorface);
    }

    //重建人脸
    for (int num = min(10, eigenvectors.cols); num < min(300, eigenvectors.cols); num+=15)
    {
        Mat evs = Mat(eigenvectors, Range::all(), Range(0, num));
        Mat projection = LDA::subspaceProject(evs, mean, image[0].reshape(1, 1));
        Mat reconstruction= LDA::subspaceReconstruct(evs, mean, projection);

        Mat result = reconstruction.reshape(1, height);
        reconstruction = normal(result, reconstruction);
        char* winTitle = new char[128];
        sprintf(winTitle, "recon_face_%d", num);
        imshow(winTitle, reconstruction);
    }
    
    waitKey(0);
}

二、FisherFace（LDA线性判别分析）

1、理论介绍

http://blog.csdn.net/feirose/article/details/39552997

 

2、流程

3、PCA和LDA的对比

 

4、代码（与特征脸代码几乎一致）

此处只列出修改部分

55行模型训练    Ptr<BasicFaceRecognizer>model = createFisherFaceRecognizer();

72行显示特征脸  for (int i = 0; i < min(16,eigenvectors.cols); i++)
                   Mat ev = eigenvectors.col(i).clone();

86行重建人脸    for (int num = 0; num < min(16, eigenvectors.cols); num++)

三、LBPH

 1、原理介绍

大家可以参考http://blog.csdn.net/xiaomaishiwoa/article/details/46640377

二、流程

3、代码

#include <opencv2/opencv.hpp>
#include <opencv2/face.hpp>

using namespace cv;
using namespace cv::face;
using namespace std;

//对原图归一化
Mat normal(Mat src, Mat dst) {
    if (src.channels() == 1)//若原图单通道
        normalize(src, dst, 0, 255, NORM_MINMAX, CV_8UC1);
    else //否则，原图三通道
        normalize(src, dst, 0, 255, NORM_MINMAX, CV_8UC3);
    return dst;
}

void main() {
    //读取文件，转换为数据流
    string filename = string("at.txt");
    ifstream file(filename.c_str(), ifstream::in);
    if (!file)
        cout << "error" << endl;

    string line, path, classlabel;
    vector<Mat>image;
    vector<int>labels;
    char separator = ';';
    while (getline(file,line))
    {
        stringstream liness(line);
        getline(liness, path, separator);
        getline(liness, classlabel);
        if (!path.empty()&&!classlabel.empty())
        {
            //cout << "path:" << path<< endl;
            image.push_back(imread(path, 0));
            labels.push_back(atoi(classlabel.c_str()));
        }
    }

    if (image.size() < 1 || labels.size() < 1)
        cout << "invalid image path..." << endl;

    int height = image[0].rows;
    int width = image[0].cols;
    //cout << "height:" << height << "," << width<<endl;

    //最后一个人为测试样本
    Mat testSample = image[image.size() - 1];
    int testLabel = labels[labels.size() - 1];
    image.pop_back();
    labels.pop_back();

    //训练
    Ptr<LBPHFaceRecognizer>model = createLBPHFaceRecognizer();
    model->train(image, labels);

    //识别
    int predictLabel = model->predict(testSample);
    cout << "actual label:" << testLabel << ",predict label:" << predictLabel << endl;

    //打印参数
    int radius = model->getRadius();    //中心像素点到周围像素点的距离
    int neibs = model->getNeighbors();    //周围像素点的个数
    int grad_x = model->getGridX();        //将一张图片在x方向分成几块
    int grad_y = model->getGridY();        //将一张图片在y方向分成几块
    double t = model->getThreshold();    //相似度阈值    
    cout << "radius:" << radius << endl;
    cout << "neibs:" << neibs << endl;
    cout << "grad_x:" << grad_x << endl;
    cout << "grad_y:" << grad_y << endl;
    cout << "threshold:" << t<<endl;
    
    waitKey(0);
}