%%% PCA-based Face Recognizer
%%% Author : Berk Gokberk
%%% Inputs : 
%%%   Training and test set images should be present in 'train_images\' and
%%%   'test_images\' directories


clear all;

dir_train = 'train_images\';
dir_test  = 'test_images\';


im_res = [0.2]; %%% image resolution parameter : 
                %%% original image resolution is 92 x 112.
                %%% all images will be resized to :
                %%% round(92 x im_res) x round(112 x im_res)

%%% calculate maximum dimensionality
curdir = cd;
cd(dir_train);
files = feval('dir','*.BMP');
[rf,cf] = size(files);
cd(curdir);
I = imread([dir_train files(1).name]);
I = imresize(I, im_res);
[I_row,I_col] = size(I);
PCA_DIM_MAX = I_row * I_col;
fprintf('Image Resolution : [%d x %d] = %d \n', I_row, I_col, I_row * I_col);


recognition_accuracy = [];

%%% Determine which PCA dimensionalities to be used
%PCA_DIM_VEC = [10:10:PCA_DIM_MAX];
PCA_DIM_VEC = [2:2:100];

for PCA_DIM = PCA_DIM_VEC

    %%% TRAINING PHASE
    %%% read training set
    curdir = cd;
    cd(dir_train);
    files = feval('dir','*.BMP');
    [rf,cf] = size(files);
    cd(curdir);

    %%% Initialize training feature matrix and class/inst label matrix
    M_train_data   = [];
    M_train_labels = [];

    fprintf('Training face recognizer...');
    for i=1:rf

        %%% read image
        I = imread([dir_train files(i).name]);

        %%% resize image
        I = imresize(I, im_res);

        %%% convert image to vector
        [I_row,I_col] = size(I);
        I_vector = double(reshape(I', 1, I_row*I_col));

        %%% add image vector to the training data matrix as a row vector
        M_train_data = [M_train_data; I_vector];
        %%% add class/instance label to the label matrix
        class_label = str2num(files(i).name(1:2));
        inst_label  = str2num(files(i).name(3));
        M_train_labels = [M_train_labels; [class_label inst_label]];
    end
    [r_M_train, c_M_train] = size(M_train_data);
    
    
    %%% calculate PCA transformation matrix : tmatrix
    %%% calculate the feature vectors of training images  :
    %%% M_train_data_PCA
    %%% Here, write your own code or function!!!
    [M_train_data_PCA, tmatrix ] = my_pca_algorithm(M_train_data, PCA_DIM);
    
    
    
    %%% TESTING PHASE

    fprintf('Testing face recognizer....\n');

    %%% read test images
    curdir = cd;
    cd(dir_test);
    files = feval('dir','*.BMP');
    [rf,cf] = size(files);
    cd(curdir);

    Test_Similarity = [];
    for i=1:rf
        I_test_image = imread([dir_test files(i).name]);

        %%% resize image
        I_test_image = imresize(I_test_image, im_res);

        test_class_label = str2num(files(i).name(1:2));
        test_inst_label  = str2num(files(i).name(3));
        [I_row, I_col] = size(I_test_image);


        %%% convert test image to vector
        I_test_vector = double(reshape(I_test_image', 1, I_row*I_col));
        
        
        %%% compute feature vector of the test image
        %%% write your own code here
        test_vector  = my_own_code;

        [rtmat, ctmat] = size(tmatrix);

        %%% compute the similarity of each training set image to the test image

        %%% Initialize similarity vector
        sim_vec = zeros(r_M_train, 3);
        for j = 1:r_M_train

            %%% calculate the similarity between the training vector and the
            %%% test vector
            train_vector = M_train_data_PCA(j, :);
            similarity   = sqrt(sum((train_vector - test_vector).^2)); %%% Euclidean distance
            sim_vec(j, :) = [similarity M_train_labels(j,:)];
        end
        %%% sort similarity vector according to similarities
        sorted_similarity = sortrows(sim_vec, 1);
        Test_Similarity(i).Test_Class = test_class_label;
        Test_Similarity(i).Test_Inst  = test_inst_label;
        Test_Similarity(i).Similarity = sorted_similarity;
        Test_Similarity(i).Filename   = files(i).name;
    end

    %%% Output classification performance
    [rTest, cTest] = size(Test_Similarity);
    rank = 1;

    %%% calculate recognition performance for a rank-n scenario
    error_count = 0;
    for i = 1:cTest
        nearest_training_class_labels = Test_Similarity(i).Similarity(1:rank, 2);
        if sum(nearest_training_class_labels == Test_Similarity(i).Test_Class) == 0
            error_count = error_count + 1;
        end
    end
    classification_accuracy  = (1 - (error_count / r_M_train)) * 100;
    recognition_accuracy = [recognition_accuracy; [PCA_DIM classification_accuracy]];
    fprintf('Rank-1 Recognition Accuracy: %5.2f\n\n', classification_accuracy);
end

figure;
plot(recognition_accuracy(:,1),recognition_accuracy(:,2), 'r.-')
axis([min(PCA_DIM_VEC) max(PCA_DIM_VEC) 0 100]);
title_str = sprintf('Recognition Accuracies for different PCA dimensionalities');
title(title_str);
xlabel('PCA dimensionality');
ylabel('Recognition Rate');
drawnow;