Task5.md

Task 5

Face Detection and Recognition

Face Detection

Image detectFace(Image &inputImage);

function parameters:

• inputImg: reference to gray Image

function return: return an image with rectangles drawn on the faces in the input image

steps:

• uses cv::CascadeClassifier to load haarcascade_frontalface_default.xml model
• convert the inputImage to cv::Mat
• uses cv::CascadeClassifier.detectMultiScale to get faces in the image
• then draw the faces on the image with Image drawFaces(cv::Mat img, std::vector<cv::Rect> faces)

Result:

Face Recognition

Steps

Training

• loading the images and convert each image to vector 1d and construct the matrix of all images

std::vector<std::string> loadImgsDataset(std::string DirPath, std::vector<std::vector<float>> &Dataset)

• get zero mean of each feature(by getting avg of each row and subtract this avg from that row)

vector<vector<float>>  GetCenteredImgs(vector<vector<float>> TrainingDataset)

• get the vector of coefficients of each image
  vector<vector<float>> GetEigenVectorsOfUpperCorr(vector<vector<float>>& TrainingDataset){
  • get covariance matrix in lower scale with size of (num Of imgs x num Of imgs) vector<vector<float>> GetCovMatrix(vector<vector<float>>& TrainingDataset)
  • get eigenvectors and eigenvalues of this covariance matrix std::vector<std::pair<std::vector<float>, float>> egienVectorsValues(const std::vector<std::vector<float>> &CovarMatrix)
  • get the eigenvectors of upper covariance matrix vector<vector<float>> GetEigenVectorsOfUpperCorr(vector<vector<float>>& TrainingDataset)
  • then get the coefficient of each image by multiplying each vector img by the matrix of above eigenvectors vector<vector<float>> getImagesCoeff(vector<vector<float>>& TrainingDataset, vector<vector<float>> & EigenVectors)

we store the eigenvectors and image coefficient to use them in testing phase.

Testing

• loading the coefficient of training images and the eigenvectors

ReadFileToVector("../Coefficient_Matrix.txt", CoffMat);
ReadFileToVector("../Eigen_Vectors_Matrix.txt", EigenVectorsUpper);

• then predict the image pair<int, float> PredictImg(vector<float> testImg, vector<vector<float>> EigenVectors, vector<vector<float>> ImgCoffMat)
  • get the coeff of test image vector<float> GetImgCoff(vector<vector<float>> EigenVectors , vector<float> ImgVector)
  • get the matched image by minimizing the error between test image coeff and train images coeffs pair<int, float> GetSimilarImg(vector<vector<float>> ImgsCoffMatrix, vector<float> TestImgCoeff) Results
    

Test Data Set

• we test all the images using function in face_recognition.h:

float TestImages()

steps:

• we test using all the test images
• we get the average distance between the test image and all the images of each class
• get the log of the distances
• use one vs all strategy :
  • for each class get the average distance of all other classes
  • the similarity is 1 - ( distance of the class / (distance of class + distance of other class )
  • the class is of the image is the class with highest similarity
• after selecting the class we get the most similar image of this class
• we write the result data into text file "roc-data.txt"