2018 CSE3/4 VIS Assignment: Eigenface Techniques for Face Recognition

This assignment explores the application of eigenface techniques for face recognition. The solution details the implementation using MATLAB, including image resizing, determining parameters (K1 and K2), and analyzing classifier output. The process involves training and testing datasets, with experiments conducted on student images to evaluate the system's ability to recognize faces under different moods. The report presents the results in tables and figures, discussing system performance, including the impact of neural networks, nearest neighbor techniques, and the effect of varying image sizes. The provided MATLAB code demonstrates the steps involved in face detection, feature extraction, and classification, along with the analysis of the results. The student's solution includes the steps of reading the images, resizing the images, determining the values for K1 and K2 and the results from the classifier output and finally the discussion of the results and the system performance.

2018
UNIVERSITY AFFILIATION
2018 CSE3/4 VIS VISUAL INFORMATION SYSTEMS ASSIGNMENT
STUDENT NAME

TABLE OF CONTENTS
BASIC OVERVIEW.......................................................................................................................2
Eigenface techniques for face recognition...................................................................................2
RESULTS AND OBSERVATIONS: MATLAB IMPLEMENTATION.......................................3
PROCEDURE:.............................................................................................................................3
Experiment and results.................................................................................................................4
Part 1: Resizing the images stored in the training dataset and test dataset..................................4
Part 2: Determining K1 and K2...................................................................................................6
Part 3: Training and Test dataset results from classifier output...................................................7
DISCUSSION..................................................................................................................................7
System performance.....................................................................................................................7
REFERENCES................................................................................................................................9
1

BASIC OVERVIEW
Eigenface techniques for face recognition
The modern security systems incorporate the use of facial recognition as a way of enabling
authorization. This can be achieved in personal identification for instance, at the immigration
office or border points, for human computer interactions such as phone unlock using facial
recognition, and major security systems for instance, in offices or large residential complex or
penthouses (Navarrete & Ruiz-del, 2002). Every human being has very complex,
multidimensional, and meaningful facial unique attributes that differentiate them from others.
This makes the facial recognition process more difficult to implement in security systems. Some
of the local features that the recognition process focusses on are the eyes, nose, and mouth before
extracting the feature of the whole face (Charalampos & Ilias, 2010).
There are a number of approaches defined to enable the facial recognition by systems. For
instance, using artificial intelligence, one can create neural networks and self-organizing maps,
SOMs, where the user trains some images and uses test images to test the recognition. Other
alternatives to this approach are content-based image retrieval, principal component analysis and
the relevance feedback (Turk & Pentland). There are three stages performed in the process of
facial recognition such as the face location detection, feature extraction, and facial image
classification. The face recognition is done using eigenface algorithm (Emad, Tamer, &
AbdelMonem). The face images are projected into a feature space that best encodes the variation
among the known face images. The face space is well expounded by the eigenfaces, which are
the eigenvectors of the set of faces. The eigenface algorithm computes the average face, v. The
algorithm collects the difference between training images and the average face (Ruiz-del-Solar &
and Navarrete, 2002). The differences are saved in a matrix where M is the number of pixels and
N is the number of stored or trained images. The algorithm for eigenfaces is denoted by the
equation,
A= [ u1
1−v ,… , un
1 −v , … ,u1
p−v , … , un
p−v ]
The eigenvectors of the covariance matrix C are used to give the final eigenfaces. This is done
using powerful tools with a stable runtime such as MATLAB R2017a. Therefore,
C= A AT
2

There are N-1 meaningful eigenvectors, when the number of data points is smaller than the
dimensions. To get a faster response on the value of the eigenvectors of C,
L= AT A
The training face images and the new face images are represented as a linear combination of the
eigenfaces. For instance, a face image, u, can be represented as,
u=∑
i
ai ϕi
The eigenvectors are usually orthogonal to the eigenvalues such that,
ai=uT ϕi
The PCA seeks directions that are efficient for the representation of the data and seeks to
maximize the total scatter. The PCA reduces the dimension of the data and speeds up the
computational time. The time taken to perform facial recognition is important especially in
implementation in the actual environment such that the systems should use the least time to
detect a face (Tolba, El-Baz, & El-Harby, 2006). It should be close to real-time.
RESULTS AND OBSERVATIONS: MATLAB IMPLEMENTATION
The task aims at using the training and testing data to identify and extract face images from the
image saved. This is done when the algorithm pulls similar images from the database with a set
of 30 images from the training data and 20 images from the test data.
PROCEDURE:
(i) The training set is obtained as the 30 images and the test set has 20 images all defined
in set folders. The algorithm is loaded and the eigenfaces are calculated using the
PCA projections. These projections define the eigenspace.
(ii) The new face is checked by using the test data and the weight of the connections or
links is computed. The system, at this point, determines if the image used is a face.
When the algorithm identifies the new image as a face, the weight pattern is grouped
under the known or unknown (Hongliang, Qingshan, Xiaoou, & Hanqing, 2005).
3

(iii) In practice, the grouping would be used to grant or deny authorization or access to a
building. The image should match an existing image in the database for it to be
successfully recognized (Belhumeur, Hespanha, & Kriegman).
(iv) The learning process during training seeks to pick the unknown faces and incorporate
them in the database. For instance, during the registration of a new staff member in a
company, the system learns and it adds the unknown face to its database alongside
some description to grant or deny access to the given face (Chellappa, Wilson, &
Sirohey).
Experiment and results.
The data is obtained from the lab experiments with student images. The training set contains 10
students with 3 moods each. The resulting matrix is 10x3 images while the test set contains 10x2
images where the 10 test subjects have 2 moods each. The designer creates a face database, the
system generates the eigenfaces database and creates the top eigenfaces (Teofilo, Rogerio, &
Roberto, 2000). The test image is used to show if the system can recognize the face from the
already setup database.
(i) The first attempt seeks to classify the average coefficient for each person such that the
new face is compared to the closest average. The recognition accuracy increases with
more learning procedures.
(ii) The eigenfaces are useful during training to set the average value of recognition but
do not help much afterwards.
Part 1: Resizing the images stored in the training dataset and test dataset
%% Part 1: Resizing and Reading Images
% Subject1-Dhara
dhara=imread('D:\ONLINE WORK\725331\assignment\Training\Dhara_Happy.jpg','jpg');
dhara=rgb2gray(dhara);
dhara=imresize(dhara,[N N1] );
figure(1),imshow(dhara,'Initialmagnification','fit');
title('dhara')
% Subject2-Diksha
diksha=imread('D:\ONLINE WORK\725331\assignment\Training\Diksha_Happy.jpg','jpg');
diksha=rgb2gray(diksha);
diksha=imresize(diksha,[N N1] );
figure(2),imshow(diksha,'Initialmagnification','fit');
4

title('dhara')
% Subject3-Eric
Eric=imread('D:\ONLINE WORK\725331\assignment\Training\Eric_Happy.jpg','jpg');
Eric=rgb2gray(Eric);
Eric=imresize(Eric,[N N1] );
figure(3),imshow(Eric,'Initialmagnification','fit');
title('Eric')
% Subject4-Gauta
Gauta=imread('D:\ONLINE WORK\725331\assignment\Training\Gautam_Happy.jpg','jpg');
Gauta=rgb2gray(Gauta);
Gauta=imresize(Gauta,[N N1] );
figure(4),imshow(Gauta,'Initialmagnification','fit');
title('Gauta')
% Subject5-Ghaida
Ghaida=imread('D:\ONLINE WORK\725331\assignment\Training\Ghaida_Happy.jpg','jpg');
Ghaida=rgb2gray(Ghaida);
Ghaida=imresize(Ghaida,[N N1] );
figure(5),imshow(Ghaida,'Initialmagnification','fit');
title('Ghaida')
% Subject6-Haoyang
Haoyang=imread('D:\ONLINE WORK\725331\assignment\Training\Haoyang_Happy.jpg','jpg');
Haoyang=rgb2gray(Haoyang);
Haoyang=imresize(Haoyang,[N N1] );
figure(6),imshow(Haoyang,'Initialmagnification','fit');
title('Haoyang')
% Subject7-Huipei
Huipei=imread('D:\ONLINE WORK\725331\assignment\Training\Huipei_Happy.jpg','jpg');
Huipei=rgb2gray(Huipei);
Huipei=imresize(Huipei,[N N1] );
figure(7),imshow(Huipei,'Initialmagnification','fit');
title('Huipei')
% Subject8-Hung
Hung=imread('D:\ONLINE WORK\725331\assignment\Training\Hung_Happy.jpg','jpg');
Hung=rgb2gray(Hung);
Hung=imresize(Hung,[N N1] );
figure(8),imshow(Hung,'Initialmagnification','fit');
title('Hung')
5

% Subject9-James
James=imread('D:\ONLINE WORK\725331\assignment\Training\James_D_Happy.jpg','jpg');
James=rgb2gray(James);
James=imresize(James,[N N1] );
figure(9),imshow(James,'Initialmagnification','fit');
title('James')
% Subject10-Neel
Neel=imread('D:\ONLINE WORK\725331\assignment\Training\Neel_Happy.jpeg','jpg');
Neel=rgb2gray(Neel);
Neel=imresize(Neel,[N N1] );
figure(10),imshow(Neel,'Initialmagnification','fit');
title('Neel')
Happy Students - Z Average
Average
Part 2: Determining K1 and K2
Table 1 Results for the test dataset (K1)
6

1-NN 3-NN 5-NN
40x30 size 39.0415 36.0745 45.6723
80x60 size 78.0830 72.1490 89.3447
Average 58.5622 54.1118 67.5085
Table 2 Results for the test dataset (K2)
1-NN 3-NN 5-NN
40x30 size 43.3975 40.4305 50.0283
80x60 size 82.439 76.505 93.7007
Average 62.9182 58.4678 71.8645
Part 3: Training and Test dataset results from classifier output
eigenfaces
DISCUSSION
System performance
The best rates that can be used in a high performing system are such as,
training set → 95 %
test set → 85 %
The neural networks model is good with training the data but it takes some time to perform the
training especially if the data set is too large. When the classification is performed using the
nearest neighbor technique, it takes more classification time. When the test case is chosen as
Dhara is surprised, the algorithm links the image to the original image of Dhara as taken during
training.
7

test face Dhara
8

REFERENCES
Belhumeur, P., Hespanha, & Kriegman, D. (n.d.). Eigenfaces vs Fisherfaces. Recognition using
Class Specific Linear Proection.
Charalampos, D., & Ilias, M. (2010). A fast-mobile face Recognition System for Android OS
Based on Eigenfaces Decomposition. IFIP Advances in Information and Communication
Technology, 295-302.
Chellappa, R., Wilson, C. L., & Sirohey, C. (n.d.). Human and Machine recognition of faces: A
Survey. Proceedings of IEEE, 705-740.
Emad, B., Tamer, M., & AbdelMonem, W. A. (n.d.). A new image comparing technique for
content-based image retrieval.
Hongliang, i., Qingshan, L., Xiaoou, T., & Hanqing, L. (2005). Learning Local Descriptors for
Face Detection: Multimedia and Expo. IEEE International Conference on 06-06 2005;
ICME 2005, 928-931.
Navarrete, P., & Ruiz-del, S. (2002). Interactive Face Retrieval using Self-Organizing Maps.
International Joint Conference on Neural Networks-IJCNN 2002, 12-17.
Ruiz-del-Solar, J., & and Navarrete, P. (2002). Towards a Generalized Eigenspace-based Face
Recognition Framework. 4th Int. Workshop on Statistical Techniques in Pattern
Recognition, 6-9.
Teofilo, E. d., Rogerio, S. F., & Roberto, C. M. (2000). First steps toward performance
assessment of representation for Face Recognition Lecture notes in Artificial Intelligence.
Eigenfaces versus Eigeneyes, 197-206.
Tolba, A. S., El-Baz, A. H., & El-Harby, A. A. (2006). Face Recognition: A literature Review.
Internation Journal of Signal Processing, 2-5.
Turk, M., & Pentland, A. (n.d.). Eigenfaces for Recognition. Journal of Cognitive Neuroscience,
71-86.
9

APPENDIX
%% Student Registration Details
% Group Members and their respective IDs
close all
clc
N=40; %size of the image
N1=30;
M=10; % 4 student faces
%% Part 1: Resizing and Reading Images
% Subject1-Dhara
dhara=imread('D:\ONLINE WORK\725331\assignment\Training\
Dhara_Happy.jpg','jpg');
dhara=rgb2gray(dhara);
dhara=imresize(dhara,[N N1] );
figure(1),imshow(dhara,'Initialmagnification','fit');
title('dhara')
% Subject2-Diksha
diksha=imread('D:\ONLINE WORK\725331\assignment\Training\
Diksha_Happy.jpg','jpg');
diksha=rgb2gray(diksha);
diksha=imresize(diksha,[N N1] );
figure(2),imshow(diksha,'Initialmagnification','fit');
title('dhara')
% Subject3-Eric
Eric=imread('D:\ONLINE WORK\725331\assignment\Training\Eric_Happy.jpg','jpg');
Eric=rgb2gray(Eric);
Eric=imresize(Eric,[N N1] );
figure(3),imshow(Eric,'Initialmagnification','fit');
title('Eric')
% Subject4-Gauta
Gauta=imread('D:\ONLINE WORK\725331\assignment\Training\
Gautam_Happy.jpg','jpg');
Gauta=rgb2gray(Gauta);
Gauta=imresize(Gauta,[N N1] );
figure(4),imshow(Gauta,'Initialmagnification','fit');
title('Gauta')
% Subject5-Ghaida
Ghaida=imread('D:\ONLINE WORK\725331\assignment\Training\
Ghaida_Happy.jpg','jpg');
Ghaida=rgb2gray(Ghaida);
Ghaida=imresize(Ghaida,[N N1] );
figure(5),imshow(Ghaida,'Initialmagnification','fit');
title('Ghaida')
10

% Subject6-Haoyang
Haoyang=imread('D:\ONLINE WORK\725331\assignment\Training\
Haoyang_Happy.jpg','jpg');
Haoyang=rgb2gray(Haoyang);
Haoyang=imresize(Haoyang,[N N1] );
figure(6),imshow(Haoyang,'Initialmagnification','fit');
title('Haoyang')
% Subject7-Huipei
Huipei=imread('D:\ONLINE WORK\725331\assignment\Training\
Huipei_Happy.jpg','jpg');
Huipei=rgb2gray(Huipei);
Huipei=imresize(Huipei,[N N1] );
figure(7),imshow(Huipei,'Initialmagnification','fit');
title('Huipei')
% Subject8-Hung
Hung=imread('D:\ONLINE WORK\725331\assignment\Training\Hung_Happy.jpg','jpg');
Hung=rgb2gray(Hung);
Hung=imresize(Hung,[N N1] );
figure(8),imshow(Hung,'Initialmagnification','fit');
title('Hung')
% Subject9-James
James=imread('D:\ONLINE WORK\725331\assignment\Training\
James_D_Happy.jpg','jpg');
James=rgb2gray(James);
James=imresize(James,[N N1] );
figure(9),imshow(James,'Initialmagnification','fit');
title('James')
% Subject10-Neel
Neel=imread('D:\ONLINE WORK\725331\assignment\Training\
Neel_Happy.jpeg','jpg');
Neel=rgb2gray(Neel);
Neel=imresize(Neel,[N N1] );
figure(10),imshow(Neel,'Initialmagnification','fit');
title('Neel')
%% reading test faces- 5 sleepy students
% Subject1-Haoyangtst
Haoyangtst=imread('D:\ONLINE WORK\725331\assignment\Testing\
haoyang_sleepy.jpg','jpg');
Haoyangtst=rgb2gray(Haoyangtst);
Haoyangtst=imresize(Haoyangtst,[N N1] );
figure(11),imshow(Haoyangtst,'Initialmagnification','fit');
title('Haoyang_Test')
% Subject2-Huipeitst
11