public LaplacianFaceRecognizer(List <Matrix> trainingSet, List <String> labels, int numOfComponents)
{
this.numOfComponents = numOfComponents;
this.imgRows = trainingSet[0].RowDimension;
int n = trainingSet.Count; // sample size
HashSet <String> tempSet = new HashSet <String>(labels);
int c = tempSet.Count; // class size
// process in PCA
EigenFaceRecognizer pca = new EigenFaceRecognizer(trainingSet, labels, numOfComponents);
//construct the nearest graph
Matrix S = constructGraph(pca.getWeights());
Matrix D = constructD(S);
Matrix L = D.Subtract(S);
//reconstruct the trainingSet into required X;
Matrix X = constructTrainingMatrix(pca.getWeights());
Matrix XLXT = X.Multiply(L).Multiply(X.Transpose());
Matrix XDXT = X.Multiply(D).Multiply(X.Transpose());
//calculate the eignevalues and eigenvectors of (XDXT)^-1 * (XLXT)
XDXT.Inverse();
Matrix targetForEigen = XDXT.Inverse().Multiply(XLXT);
EigenvalueDecomposition feature = targetForEigen.Eigen();
double[] eigenValues = feature.RealEigenvalues;
int[] indexOfChosenEigenValues = this.getIndOfHigherEV(eigenValues, eigenValues.Length);//higher eigen values
Matrix eigenVectors = feature.GetV();
Matrix selectedEigenVectors = eigenVectors.GetMatrix(0, eigenVectors.RowDimension - 1, indexOfChosenEigenValues);
this.weightMatrix = pca.getWeightMatrix().Multiply(selectedEigenVectors);
//Construct projectedTrainingMatrix
this.weights = new List <ProjectMatrix>();
for (int i = 0; i < trainingSet.Count(); i++)
{
ProjectMatrix pm = new ProjectMatrix(this.weightMatrix.Transpose().Multiply(trainingSet[i].Subtract(pca.getMeanMatrix())),
labels[i]);
this.weights.Add(pm);
}
this.meanMatrix = pca.getMeanMatrix();
}
public FisherFaceRecognizer(List <Matrix> trainingSet, List <String> labels, int numOfComponents)
{
int n = trainingSet.Count; // sample size
HashSet <String> tempSet = new HashSet <String>(labels);
int c = tempSet.Count; // class size
// process in PCA
EigenFaceRecognizer pca = new EigenFaceRecognizer(trainingSet, labels, numOfComponents);
//construct the nearest neighbor graph
Matrix S = constructNearestNeighborGraph(pca.getProjectSet());
Matrix D = constructD(S);
Matrix L = D.Subtract(S);
//reconstruct the trainingSet into required X;
Matrix X = constructTrainingMatrix(pca.getProjectSet());
Matrix XLXT = X.Multiply(L).Multiply(X.Transpose());
Matrix XDXT = X.Multiply(D).Multiply(X.Transpose());
//calculate the eignevalues and eigenvectors of (XDXT)^-1 * (XLXT)
Matrix targetForEigen = XDXT.Inverse().Multiply(XLXT);
EigenvalueDecomposition feature = targetForEigen.Eigen();
double[] d = feature.RealEigenvalues;
//assert d.length >= c - 1 :"Ensure that the number of eigenvalues is larger than c - 1";///
int[] indexes = getIndOfHigherEV(d, d.Length);
Matrix eigenVectors = feature.GetV();
Matrix selectedEigenVectors = eigenVectors.GetMatrix(0, eigenVectors.RowDimension - 1, indexes);
this.weightMatrix = pca.getWeightMatrix().Multiply(selectedEigenVectors);
//Construct projectedTrainingMatrix
this.projectSet = new List <ProjectMatrix>();
for (int i = 0; i < trainingSet.Count(); i++)
{
ProjectMatrix ptm = new ProjectMatrix(this.weightMatrix.Transpose().Multiply(trainingSet[i].Subtract(pca.getMeanMatrix())), labels[i]);
this.projectSet.Add(ptm);
}
this.meanMatrix = pca.getMeanMatrix();
}
public LDA(List <Matrix> trainingSet, List <String> labels,
int numOfComponents)
{
int n = trainingSet.Count(); // sample size
HashSet <string> tempSet = new HashSet <string>(labels);
int c = tempSet.Count(); // class size
/// deh mfrod used for debugging issues, so fakes for nw //////////////////////////
//assert numOfComponents >= n - c : "the input components is smaller than n - c!";
//assert n >= 2 * c : "n is smaller than 2c!";
// process in PCA
EigenFaceRecognizer pca = new EigenFaceRecognizer(trainingSet, labels, n - c);
// classify
Matrix meanTotal = new Matrix(n - c, 1);
Dictionary <string, List <Matrix> > map = new Dictionary <string, List <Matrix> >();
List <ProjectMatrix> pcaTrain = pca.getProjectSet();
for (int i = 0; i < pcaTrain.Count(); i++)
{
string key = pcaTrain[i].getLabel();
meanTotal.AddEquals(pcaTrain[i].getImgMat());
if (!map.ContainsKey(key))
{
List <Matrix> temp = new List <Matrix>();
temp.Add(pcaTrain[i].getImgMat());
map.Add(key, temp);
}
else
{
List <Matrix> temp = map[key];
temp.Add(pcaTrain[i].getImgMat());
map[key] = temp;
}
}
meanTotal.Multiply((double)1 / n);
// calculate Sw, Sb
Matrix Sw = new Matrix(n - c, n - c);
Matrix Sb = new Matrix(n - c, n - c);
/*** !!! **/
tempSet = new HashSet <string>(map.Keys);
/*** !!! **/
foreach (string s in tempSet)
{
//iterator<string> it = tempSet.iterator();
//while (it.hasNext()) {
//String s = (String)it.next();
List <Matrix> matrixWithinThatClass = map[s];
Matrix meanOfCurrentClass = getMean(matrixWithinThatClass);
for (int i = 0; i < matrixWithinThatClass.Count(); i++)
{
Matrix temp1 = matrixWithinThatClass[i].Subtract(meanOfCurrentClass);
temp1 = temp1.Multiply(temp1.Transpose());
Sw.AddEquals(temp1);
}
Matrix temp = meanOfCurrentClass.Subtract(meanTotal);
temp = temp.Multiply(temp.Transpose()).Multiply(matrixWithinThatClass.Count());
Sb.AddEquals(temp);
}
// calculate the eigenvalues and vectors of Sw^-1 * Sb
Matrix targetForEigen = Sw.Inverse().Multiply(Sb);
EigenvalueDecomposition feature = targetForEigen.Eigen();
double[] d = feature.RealEigenvalues;
//assert d.length >= c - 1 : "Ensure that the number of eigenvalues is larger than c - 1";
int[] indexes = getIndOfHigherEV(d, c - 1);
Matrix eigenVectors = feature.GetV();
Matrix selectedEigenVectors = eigenVectors.GetMatrix(0, eigenVectors.RowDimension - 1, indexes);
this.weightMatrix = pca.getWeightMatrix().Multiply(selectedEigenVectors);
// Construct projectedTrainingMatrix
this.projectSet = new List <ProjectMatrix>();
for (int i = 0; i < trainingSet.Count(); i++)
{
ProjectMatrix ptm = new ProjectMatrix(this.weightMatrix
.Transpose()
.Multiply(trainingSet[i].Subtract(pca.getMeanMatrix())),
labels[i]);
this.projectSet.Add(ptm);
}
this.meanMatrix = pca.getMeanMatrix();
}
static void train(int metricType, int recognizer, int noOfComponents, int trainNums, int knn_k)
{
//set trainSet and testSet
Dictionary <String, List <int> > trainMap = new Dictionary <string, List <int> >();
Dictionary <String, List <int> > testMap = new Dictionary <string, List <int> >();
for (int i = 1; i <= 5; i++)
{
String label = "s" + i;
List <int> train = generateTrainNums(trainNums);
List <int> test = generateTestNums(train);
trainMap[label] = train;
testMap[label] = test;
}
//trainingSet & respective labels
List <Matrix> trainingSet = new List <Matrix>();
List <String> labels = new List <String>();
HashSet <String> labelSet = new HashSet <string>(trainMap.Keys);
foreach (String label in labelSet)
{
List <int> cases = trainMap[label];
for (int i = 0; i < cases.Count; i++)
{
String filePath = "E:/faces/" + label + "/" + cases[i] + ".bmp";
Matrix temp;
try
{
temp = ImageGrapper.GetMatrixGivenPath(filePath);
if (recognizer != 0)
{
temp = convertTo1D(temp);
}
trainingSet.Add(temp);
labels.Add(label);
} catch (Exception e) {}
}
}
FaceRecognizer fr = null;
if (recognizer == 0)
{
fr = new ModularFaceRecognitionAlgorithms(trainingSet, labels, 200, 16, trainNums);
}
else if (recognizer == 1)
{
fr = new EigenFaceRecognizer(trainingSet, labels, 40);
}
else if (recognizer == 2)
{
fr = new LinearDescriminantAnalysis(trainingSet, labels);
}
else if (recognizer == 3)
{
fr = new LaplacianFaceRecognizer(trainingSet, labels, 20);
}
fr.saveTrainingData(@"E:\trainLPP.txt");
}
public LinearDescriminantAnalysis(List <Matrix> trainingSet, List <String> labels)
{
this.trainingSet = trainingSet;
this.imgRows = trainingSet[0].RowDimension;
int n = trainingSet.Count(); // sample size
HashSet <string> uniqueLabels = new HashSet <string>(labels);
int c = uniqueLabels.Count(); // class size
this.numOfComponents = c - 1;
EigenFaceRecognizer pca = new EigenFaceRecognizer(trainingSet, labels, n - c);
Matrix meanTotal = new Matrix(n - c, 1);
Dictionary <string, List <Matrix> > dic = new Dictionary <string, List <Matrix> >();
List <ProjectMatrix> pcaWeights = pca.getWeights();
for (int i = 0; i < pcaWeights.Count(); i++)
{
string key = pcaWeights[i].getLabel();
meanTotal.AddEquals(pcaWeights[i].getImgMat());
if (!dic.ContainsKey(key))
{
List <Matrix> temp = new List <Matrix>();
temp.Add(pcaWeights[i].getImgMat());
dic.Add(key, temp);
}
else
{
List <Matrix> temp = dic[key];
temp.Add(pcaWeights[i].getImgMat());
dic[key] = temp;
}
}
meanTotal.Multiply((double)1 / n);
// calculate Sw, Sb
Matrix Sw = new Matrix(n - c, n - c);
Matrix Sb = new Matrix(n - c, n - c);
uniqueLabels = new HashSet <string>(dic.Keys);
foreach (string s in uniqueLabels)
{
List <Matrix> matrixWithinThatClass = dic[s];
Matrix meanOfCurrentClass = getMean(matrixWithinThatClass);
for (int i = 0; i < matrixWithinThatClass.Count(); i++)
{
Matrix mat = matrixWithinThatClass[i].Subtract(meanOfCurrentClass);
mat = mat.Multiply(mat.Transpose());
Sw.AddEquals(mat);
}
Matrix temp = meanOfCurrentClass.Subtract(meanTotal);
temp = temp.Multiply(temp.Transpose()).Multiply(matrixWithinThatClass.Count());
Sb.AddEquals(temp);
}
// calculate the eigenvalues and vectors of Sw^-1 * Sb
Matrix targetForEigen = Sw.Inverse().Multiply(Sb);
EigenvalueDecomposition feature = targetForEigen.Eigen();
double[] eigenValues = feature.RealEigenvalues;
int[] indexOfChosenEigenValues = getIndOfHigherEV(eigenValues, c - 1);
Matrix eigenVectors = feature.GetV();
Matrix selectedEigenVectors = eigenVectors.GetMatrix(0, eigenVectors.RowDimension - 1, indexOfChosenEigenValues);
this.weightMatrix = pca.getWeightMatrix().Multiply(selectedEigenVectors);
// Construct weights
this.weights = new List <ProjectMatrix>();
for (int i = 0; i < trainingSet.Count(); i++)
{
ProjectMatrix ptm = new ProjectMatrix(this.weightMatrix.Transpose().Multiply(trainingSet[i].Subtract(pca.getMeanMatrix()))
, labels[i]);
this.weights.Add(ptm);
}
this.meanMatrix = pca.getMeanMatrix();
}
/*metricType:
* 0: CosineDissimilarity
* 1: L1Distance
* 2: EuclideanDistance
*
* energyPercentage:
* PCA: components = samples * energyPercentage
* LDA: components = (c-1) *energyPercentage
* LLP: components = (c-1) *energyPercentage
*
* featureExtractionMode
* 0: PCA
* 1: LDA
* 2: LLP
*
* trainNums: how many numbers in 1..10 are assigned to be training faces
* for each class, randomly generate the set
*
* knn_k: number of K for KNN algorithm
*
* */
static double test(int metricType, int noOfComponents, int faceRecognizerFlag, int trainNums, int knn_k)
{
//determine which metric is used
//metric
Metric metric = null;
if (metricType == 0)
{
metric = new CosineDissimilarity();
}
else if (metricType == 1)
{
metric = new L1Distance();
}
else if (metricType == 2)
{
metric = new EuclideanDistance();
}
//////////assert metric != null : "metricType is wrong!";////////
//set expectedComponents according to energyPercentage
//componentsRetained
// int trainingSize = trainNums * 10;
// int componentsRetained = 0;
// if(featureExtractionMode == 0)
// componentsRetained = (int) (trainingSize * energyPercentage);
// else if(featureExtractionMode == 1)
// componentsRetained = (int) ((10 -1) * energyPercentage);
// else if(featureExtractionMode == 2)
// componentsRetained = (int) ((10 -1) * energyPercentage);
//set trainSet and testSet
Dictionary <String, List <int> > trainMap = new Dictionary <string, List <int> >();
Dictionary <String, List <int> > testMap = new Dictionary <string, List <int> >();
for (int i = 1; i <= 40; i++)
{
String label = "s" + i;
List <int> train = generateTrainNums(trainNums);
List <int> test = generateTestNums(train);
trainMap[label] = train;
testMap[label] = test;
}
//trainingSet & respective labels
List <Matrix> trainingSet = new List <Matrix>();
List <String> labels = new List <String>();
HashSet <String> labelSet = new HashSet <string>(trainMap.Keys);
foreach (String label in labelSet)
{
List <int> cases = trainMap[label];
for (int i = 0; i < cases.Count; i++)
{
String filePath = "E:/faces/" + label + "/" + cases[i] + ".bmp";
Matrix temp;
try {
temp = FileManager.GetBitMapColorMatrix(filePath);
if (faceRecognizerFlag == 3)
{
trainingSet.Add(temp);
}
else
{
trainingSet.Add(convertTo1D(temp));
}
labels.Add(label);
} catch (Exception e) {
}
}
}
//testingSet & respective true labels
List <Matrix> testingSet = new List <Matrix>();
List <String> trueLabels = new List <String>();
labelSet = new HashSet <string>(trainMap.Keys);
foreach (string label in labelSet)
{
List <int> cases = testMap[label];
for (int i = 0; i < cases.Count(); i++)
{
String filePath = "E:/faces/" + label + "/" + cases[i] + ".bmp";
Matrix temp;
try
{
temp = FileManager.GetBitMapColorMatrix(filePath);
testingSet.Add(convertTo1D(temp));
trueLabels.Add(label);
}
catch (Exception e) { }
}
}
//set featureExtraction
FaceRecognizer fe = null;
if (faceRecognizerFlag == 0)
{
fe = new EigenFaceRecognizer(trainingSet, labels, noOfComponents);
}
else if (faceRecognizerFlag == 1)
{
fe = new LDA(trainingSet, labels, noOfComponents);
}
else if (faceRecognizerFlag == 2)
{
fe = new FisherFaceRecognizer(trainingSet, labels, noOfComponents);
}
else if (faceRecognizerFlag == 3)
{
fe = new ModularFaceRecognitionAlgorithms(trainingSet, labels, 200, 16, trainNums);
}
FileManager.convertMatricetoImage(fe.getWeightMatrix(), faceRecognizerFlag);
//use test cases to validate
//testingSet trueLables
List <ProjectMatrix> projectSet = fe.getProjectSet();
int accurateNum = 0;
for (int i = 0; i < testingSet.Count; i++)
{
Matrix testCase = fe.getWeightMatrix().Transpose().Multiply(testingSet[i].Subtract(fe.getMeanMatrix()));
String result = KNN.assignLabel(projectSet.ToArray(), testCase, knn_k, metric);
if (result == trueLabels[i])
{
accurateNum++;
}
}
double accuracy = accurateNum / (double)testingSet.Count;
Console.WriteLine("The accuracy is " + accuracy);
return(accuracy);
}