main_fitcecoc.m

%----------------------------------------------------------------
% File:     main_fitcecoc.m
%----------------------------------------------------------------
%
% Author:   Marek Rychlik (rychlik@arizona.edu)
% Date:     Fri Nov 22 20:02:05 2024
% Copying:  (C) Marek Rychlik, 2020. All rights reserved.
% 
%----------------------------------------------------------------
% Classification into several classes
% This script trains a facial recognition model. The model
% is saved to a .MAT file, along with necessary data to perform facial
% recognition:

targetSize = [128,128];
k=30;                                   % Number of features to consider
location = fullfile('lfw');

disp('Creating image datastore...');
imds0 = imageDatastore(location,'IncludeSubfolders',true,'LabelSource','foldernames',...
                      'ReadFcn', @(filename)imresize(im2gray(imread(filename)),targetSize));

disp('Creating subset of several persons...');
persons = {'Angelina_Jolie', 'Eduardo_Duhalde', 'Amelie_Mauresmo'}


% NOTE: Alternatively, you could pick people based on some criteria
% Below you find code that picks people with at least 10 and not more
% than 40 images
%
% tbl = countEachLabel(imds0);
% mask = tbl{:,2}>=10 & tbl{:,2}<=40;
% disp(['Number of images: ',num2str(sum(tbl{mask,2}))]);
% persons = unique(tbl{mask,1});


[lia, locb] = ismember(imds0.Labels, persons);
imds = subset(imds0, lia);

t=tiledlayout('flow');
nexttile(t);
montage(imds);

disp('Reading all images');
A = readall(imds);

B = cat(3,A{:});
D = prod(targetSize);
B = reshape(B,D,[]);

disp('Normalizing data...');
B = single(B)./256;
% NOTE: Normalization subtracts the mean pixel value
% from all pixels and divides by standard deviation. It is
% equivalent to:
%     [B,C,SD] = normalize(B, 1)
% This procedure is different from an alternative:
%     [B,C,SD] = normalize(B, 2)
% which computes the 'mean face' and subtracts it from every
% face. SD is then the l^2-norm between a face and mean face.
[B,C,SD] = normalize(B);
tic;
[U,S,V] = svd(B,'econ');
toc;

% Get an montage of eigenfaces
Eigenfaces = arrayfun(@(j)reshape((U(:,j)-min(U(:,j)))./(max(U(:,j))-min(U(:,j))),targetSize), ...
    1:size(U,2),'uni',false);

nexttile(t);
montage(Eigenfaces(1:16));
title('Top 16 Eigenfaces');
colormap(gray);

% NOTE: Rows of V are observations, columns are features.
% Observations need to be in rows.
k = min(size(V,2),k);

% Discard unnecessary data
W = S * V';                             % Transform V to weights (ala PCA)
W = W(1:k,:);                           % Keep first K weights
% NOTE: We will never again need singular values S
%S = diag(S);
%S = S(1:k);
U = U(:,1:k);                           % Keep K eigenfaces

% Find feature vectors of all images
X = W';
Y = categorical(imds.Labels, persons);

% Create colormap
cm=[1,0,0;
    0,0,1,
    0,1,0];
% Assign colors to target values
c=cm(1+mod(uint8(Y),size(cm,1)),:);

disp('Training Support Vector Machine...');
options = statset('UseParallel',true);
tic;

% You may try this, to get a more optimized model
% 'OptimizeHyperparameters','all',...

Mdl = fitcecoc(X, Y,'Verbose', 2,'Learners','svm',...
               'Options',options);
toc;

% Generate a plot in feature space using top two features
nexttile(t);
scatter3(X(:,1),X(:,2),X(:,3),50,c);
title('A top 3-predictor plot');
xlabel('x1');
ylabel('x2');
zlabel('x3');

nexttile(t);
scatter3(X(:,4),X(:,5),X(:,6),50,c);
title('A next 3-predictor plot');
xlabel('x4');
ylabel('x5');
zlabel('x6');

%[YPred,Score] = predict(Mdl,X);
[YPred,Score,Cost] = resubPredict(Mdl);

% ROC = receiver operating characteristic
% See https://en.wikipedia.org/wiki/Receiver_operating_characteristic
disp('Plotting ROC metrics...');
rm = rocmetrics(imds.Labels, Score, persons);
nexttile(t);
plot(rm);

disp('Plotting confusion matrix...')
nexttile(t);
confusionchart(Y, YPred);
title(['Number of features: ' ,num2str(k)]);

% Save the model and persons that the model recognizes.
% NOTE: An important part of the submission.
save('model','Mdl','persons','U', 'targetSize');