Content uploaded by Wael Ghazi Alnahari
Author content
All content in this area was uploaded by Wael Ghazi Alnahari on Oct 23, 2020
Content may be subject to copyright.
Kingdom of Saudi Arabia
Ministry of Education
Bisha University
Faculty of Computers and IT
Information System Department
Master in Cybersecurity
Biometric Systems
SYS 616
A Short Technical Report in
A CONVOLUTIONAL NEURAL NETWORK FOR
IRIS RECOGNITION
Supervisor(s)
Dr. Mostafa Ahmad
Assistant Professor in Computer Science Department
Author(s)
Wael Ghazi Ahmed Alnahari
Final Practical Report
April 2020
ii
TABLE OF CONTENTS
ABSTRACT ...................................................................................................................................III
INTRODUCTION ...........................................................................................................................1
1.1 Introduction to CNN ................................................................................................1
1.2 CNN Architecture ....................................................................................................2
IMPLEMENTATION OF CNN FOR BIOMETRIC RECOGNITION ..........................................3
2.1 Data Set ....................................................................................................................3
2.2 Layers .......................................................................................................................6
2.3 Training Options ......................................................................................................8
RESULTS ........................................................................................................................................9
3.1 Training Trails .........................................................................................................9
3.2 Final Accuracy .......................................................................................................16
3.3 Samples of Identification Results ..........................................................................19
APPENDIX: MATLAB CODE .....................................................................................................20
ACKNOLEDGEMENT .................................................................................................................30
REFERENCES ..............................................................................................................................31
iii
ABSTRACT
In this paper, I proposed an iris recognition system by using deep learning via convolutional
neural networks (CNN). Although CNN is used for machine learning, the recognition is achieved
by building a non-trained CNN network with multiple layers. The main objective of the code is
recognizing test pictures’ category (aka person name) with high accuracy rate after having
extracted enough features from training pictures of the same category which are obtained from a
dataset that I added to the code. I used IITD iris dataset which included 10 iris pictures for 223
people. The pictures were divided into 3 pictures for testing and 7 pictures for training. The
categories are from 001 to 223. Since the number of pictures for training is low, in order to
enhance the recognition accuracy of the network, I used five sets of layers and altered nine
parameters of sgdm training option. The code concluded an accuracy rate of 97.46% and the time
elapsed was 10 minutes and 30 seconds.
1
INTRODUCTION
1.1 Introduction to CNN
Convolutional neural networks (CNN) is a deep learning algorithm used to help machines
categorize objects. CNN is used as a computer vision detection on images. CNN is designed to
mimic human brain hence the name neural. CNN is used to treat visual inputs and determine
what the input represents based on the inputs features. CNN uses a set of multiple layers to
analyze the visual inputs and to determine what are the inputs’ distinguishing features based on
probability of repetition. CNN is useful for Artificial Intelligence (AI) such as self-driving cars
to help the cars determine what the objects appearing in their cameras represent. In general, CNN
is two parts: convolution layer and fully connected layer. The convolution layer is a single layer
used to extract a specific feature from the visual input which can be repeated as many times as
desired. The fully connected layer connects all convolution layers to one layer or representation
to combine the features.[1]
2
1.2 CNN Architecture
In order to create a CNN system, one needs a dataset to be used for training and another dataset
to be used for testing. The training dataset is categorized according to the desirable categories.
The set of data for training is usually very large compared to the test data but it is not a
requirement; however, the bigger the number of data for training the better the quality of the
result. The machine uses the training images or train dataset to find distinguishing features of
each category. The features are extracted with the help of layers. A layer in this case represents a
specific process which transforms the image into another shape, size, color, or look which is
done in some pixels that are extremely smaller than the actual image. Many layers are applied for
images of the same category which are then stored in the network of the CNN. The stored
features are based on the probability of repetition of a number of features among a category from
the training dataset. Afterward, when an image is tested, the same layers will be applied to it and
then are processed to determine which category is most similar to its features.
3
IMPLEMENTATION OF CNN FOR BIOMETRIC
RECOGNITION
2.1 Data Set
Description of the IITD Iris Image Database version 1.0
=======================================================
This iris image database mainly consists of the iris images collected from the students
and staff at IIT Delhi, India. This database has been acquired in
the Biometrics Research Laboratory during January - July 2007 using JIRIS, JPC1000,
digital CMOS camera. The acquired images were saved in bitmap format.
The database of 2240 images is acquired from 224 different users and made available
freely to the researchers. All the subjects in the database are in the
age group 14-55 years comprising of 176 males and 48 females. The resolution of these
images is 320 x 240 pixels and all these images were acquired in the
indoor environment. All the images in the database were acquired from the volunteers
who were not paid or provided any honorarium. The images were acquired
using an automated program that requires users to present their eyes in a sequence until
ten images are registered.
Organization of Database
========================
4
The acquired database is saved in 224 folders, each corresponding to 224 subjects.
Majority of images were acquired from the left eyes while the rest images
were acquired from right eye. Now the database has a label 'L' or 'R' which designates left
or right eye. There are 1288 images from 224 subject that are from
left eyes while the rest images from 211 subjects are from right eyes.Except folders 1-13,
27, 55 and 65 all other folders have five left and 5 right eye
images. (**appended on 20-04-2016**).
Usage of Database
========================
This database is only available for research and noncommercial purposes. Commercial
distribution or any act related to commercial use of this database is strictly
prohibited. Kindly acknowledge all the publicly available publications/work employing
this database with the following acknowledgment:
"Portions of the work tested on the IITD Iris Database version 1.0"
A citation to "IIT Delhi Iris Database version 1.0,
http://web.iitd.ac.in/~biometrics/Database_Iris.htm
Related Publication:
====================
Ajay Kumar and Arun Passi, "Comparison and combination of iris matchers for reliable
personal identification," Proc. CVPR 2008, Anchorage, Alaska, pp. 21-27 Jun. 2008
6
2.2 Layers
A CNN is composed of different types of layers as follows: [1]
Convolutional layer: in this layer the computer scans the visual input few pixels at a time
to determine a feature for each few pixels then continue to the next few pixels until the
whole visual input is finalized.
Pooling layer (downsampling): in this layer the features of the previous layer is
downsampled after having stored the most important features.
Fully connected layer: in this layer all the previous layers are connected together to build
a conclusion of all important features.
The used layers of our CNN are:
layers = [
imageInputLayer([x y z]);
convolution2dLayer(3,8,'Padding','same')
batchNormalizationLayer
reluLayer();
maxPooling2dLayer(5,'Stride',2)
7
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer();
averagePooling2dLayer(5,'Stride',2)
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer();
maxPooling2dLayer(5,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
batchNormalizationLayer
reluLayer();
averagePooling2dLayer(5,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
batchNormalizationLayer
reluLayer();
fullyConnectedLayer(223,'BiasLearnRateFactor',2);
softmaxLayer
classificationLayer];
8
2.3 Training Options
options = trainingOptions('sgdm', ...
'InitialLearnRate', 0.0001, ...
'ValidationData',imdsValidation, ...
'ValidationFrequency',30, ...
'Shuffle','every-epoch', ...
'MaxEpochs', 10, ...
'MiniBatchSize', 8, ...
'ValidationFrequency',50, ...
'LearnRateSchedule','piecewise', ...
'LearnRateDropFactor',0.05, ...
'LearnRateDropPeriod',60, ...
'ExecutionEnvironment','parallel',...
'Verbose', true, 'Plots','training-progress');
9
RESULTS
3.1 Training Trails
I got an accuracy of 54.56% using the following specifications for layers and options:
%% Define Network Architecture
layers = [
imageInputLayer([x y z]);
convolution2dLayer(3,8,'Padding','same')
batchNormalizationLayer
reluLayer
maxPooling2dLayer(2,'Stride',2)
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer
maxPooling2dLayer(2,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
batchNormalizationLayer
reluLayer
10
fullyConnectedLayer(size(categories,2),'BiasLearnRateFa
ctor',2);
softmaxLayer
classificationLayer];
%% Specify Training Options
options = trainingOptions('sgdm', ...
'InitialLearnRate', 0.01, ...
'ValidationData',imdsValidation, ...
'ValidationFrequency',35, ...
'Shuffle','every-epoch', ...
'MaxEpochs', 15, ...
'MiniBatchSize', 8, ...
'LearnRateSchedule','piecewise', ...
'LearnRateDropFactor',0.05, ...
'LearnRateDropPeriod',60, ...
'ExecutionEnvironment','parallel', ...
'Verbose', true, 'Plots','training-progress');
11
When I adjusted 'InitialLearnRate' to 0.001 instead of 0.01 I got an
accuracy rate of 92.83%
12
When I used the same training options as above but the maxEpoch to 6 in order to
compare the results with the previous trial and changed the layers I managed to reduce
the elapsed time to 00:10:41 with 96.41% accuracy
%% Define Network Architecture
layers = [
imageInputLayer([x y z]);
convolution2dLayer(3,8,'Padding','same')
batchNormalizationLayer
reluLayer();
maxPooling2dLayer(5,'Stride',2)
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer();
averagePooling2dLayer(5,'Stride',2)
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer();
13
maxPooling2dLayer(5,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
batchNormalizationLayer
reluLayer();
averagePooling2dLayer(5,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
batchNormalizationLayer
reluLayer();
fullyConnectedLayer(size(categories,2),'BiasLearnRateFa
ctor',2);
softmaxLayer
classificationLayer];
%% Specify Training Options
options = trainingOptions('sgdm', ...
'InitialLearnRate', 0.001, ...
'ValidationData',imdsValidation, ...
'ValidationFrequency',35, ...
14
'Shuffle','every-epoch', ...
'MaxEpochs', 6, ...
'MiniBatchSize', 8, ...
'LearnRateSchedule','piecewise', ...
'LearnRateDropFactor',0.05, ...
'LearnRateDropPeriod',60, ...
'ExecutionEnvironment','parallel', ...
'Verbose', true, 'Plots','training-progress');
When I changed 'InitialLearnRate' to 0.0001 and
'ValidationFrequency' to 35 I ended up with a lower accuracy 86.70%
15
When I changed 'InitialLearnRate' to 0.00001 I ended up with an
extremely low accuracy of 15.55%
16
3.2 Final Accuracy
Using the following layer and training options:
layers = [
imageInputLayer([x y z]);
convolution2dLayer(3,8,'Padding','same')
batchNormalizationLayer
reluLayer();
maxPooling2dLayer(5,'Stride',2)
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer();
averagePooling2dLayer(5,'Stride',2)
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer();
maxPooling2dLayer(5,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
17
batchNormalizationLayer
reluLayer();
averagePooling2dLayer(5,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
batchNormalizationLayer
reluLayer();
fullyConnectedLayer(size(categories,2),'BiasLearnRateFa
ctor',2);
softmaxLayer
classificationLayer];
%% Specify Training Options
options = trainingOptions('sgdm', ...
'InitialLearnRate', 0.001, ...
'ValidationData',imdsValidation, ...
'ValidationFrequency',50, ...
'Shuffle','every-epoch', ...
'MaxEpochs', 10, ...
18
'MiniBatchSize', 8, ...
'LearnRateSchedule','piecewise', ...
'LearnRateDropFactor',0.5, ...
'LearnRateDropPeriod',50, ...
'ExecutionEnvironment','parallel', ...
'Verbose', true, 'Plots','training-progress');
19
3.3 Samples of Identification Results
The for loop for samples was adjusted to display all mismatches then a sample of
successful matches were displayed with the same numbers of mismatches as follows:
20
APPENDIX: MATLAB CODE
CNN Code for Biometric Recognition
% Biometric Systems
% CYS616
% Wael Ghazi Ahmed Alnahari
% 440804845
% Dr. Mostafa Abdel-Halim Mostafa Ahmad
% A CONVOLUTIONAL NEURAL NETWORK FOR IRIS RECOGNITION
clc; % Clear the command window.
close all; % Close all figures (except those of imtool.)
clear; % Erase all existing variables. Or clearvars if you
want.
workspace; % Make sure the workspace panel is showing.
reset(gpuDevice); % Reset GPU memory
%% Load train Data
categories =
{'001','002','003','004','005','006','007','008','009','010
','011','012','013','014','015','016','017','018','019','02
0', ...
21
'021','022','023','024','025','026','027','028','029','030'
,'031','032','033','034','035','036','037','038','039','040
', ...
'041','042','043','044','045','046','047','048','049','050'
,'051','052','053','054','055','056','057','058','059','060
', ...
'061','062','063','064','065','066','067','068','069','070'
,'071','072','073','074','075','076','077','078','079','080
', ...
'081','082','083','084','085','086','087','088','089','090'
,'091','092','093','094','095','096','097','098','099','100
', ...
'101','102','103','104','105','106','107','108','109','110'
,'111','112','113','114','115','116','117','118','119','120
', ...
'121','122','123','124','125','126','127','128','129','130'
22
,'131','132','133','134','135','136','137','138','139','140
', ...
'141','142','143','144','145','146','147','148','149','150'
,'151','152','153','154','155','156','157','158','159','160
', ...
'161','162','163','164','165','166','167','168','169','170'
,'171','172','173','174','175','176','177','178','179','180
', ...
'181','182','183','184','185','186','187','188','189','190'
,'191','192','193','194','195','196','197','198','199','200
', ...
'201','202','203','204','205','206','207','208','209','210'
,'211','212','213','214','215','216','217','218','219','220
', ...
'221','222','223'};
imdsTrain =
imageDatastore(fullfile(pwd,"Dataset\TrainData",
23
categories),'IncludeSubfolders',true,'FileExtensions','.bmp
','LabelSource','foldernames');
num_train = size(imdsTrain.Files,1);
%% Size of first image in dataset
img = readimage(imdsTrain,1);
[x , y , z] = size(img);
%% Load Test Data
imdsValidation =
imageDatastore(fullfile(pwd,"Dataset\TestData",
categories),'IncludeSubfolders',true,'FileExtensions','.bmp
','LabelSource','foldernames');
num_test = size(imdsValidation.Files,1);
%% Calculate the number of images in each category.
labelCount = countEachLabel(imdsTrain);
%% Define Network Architecture
layers = [
imageInputLayer([x y z]);
24
convolution2dLayer(3,8,'Padding','same')
batchNormalizationLayer
reluLayer();
maxPooling2dLayer(5,'Stride',2)
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer();
averagePooling2dLayer(5,'Stride',2)
convolution2dLayer(3,16,'Padding','same')
batchNormalizationLayer
reluLayer();
maxPooling2dLayer(5,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
batchNormalizationLayer
reluLayer();
averagePooling2dLayer(5,'Stride',2)
convolution2dLayer(3,32,'Padding','same')
batchNormalizationLayer
25
reluLayer();
fullyConnectedLayer(size(categories,2),'BiasLearnRateFactor
',2);
softmaxLayer
classificationLayer];
%% Specify Training Options
options = trainingOptions('sgdm', ...
'InitialLearnRate', 0.001, ...
'ValidationData',imdsValidation, ...
'ValidationFrequency',50, ...
'Shuffle','every-epoch', ...
'MaxEpochs', 10, ...
'MiniBatchSize', 8, ...
'LearnRateSchedule','piecewise', ...
'LearnRateDropFactor',0.5, ...
'LearnRateDropPeriod',50, ...
'ExecutionEnvironment','parallel', ...
'Verbose', true, 'Plots','training-progress');
26
%% Train Network Using Training Data
[net_Wael, info] =
trainNetwork(imdsTrain,layers,options);
save net_Wael
%% Classify validation
labels = classify(net_Wael,imdsValidation);
%% *Test one at a time*
true=0;
false=0;
for i = 1:size(imdsValidation.Files,1)
im = imread(imdsValidation.Files{i});
if labels(i) == imdsValidation.Labels(i)
true=true+1;
else
false=false+1;
end
end
27
% Find the square root of the false matches to
determine the subplots
f0=floor(sqrt(false));
if f0 == sqrt(false)
f1=f0;
else
f1=f0+1;
end
% Recalbirating true and false
true=0;
false=0;
%start plotting mismatches
figure('Name','All Mismatch Pictures')
for i = 1:size(imdsValidation.Files,1)
im = imread(imdsValidation.Files{i});
if labels(i) ~= imdsValidation.Labels(i)
colorText = 'r';
false=false+1;
28
if false > f0*f1
break
else
subplot(f1,f0,false);
imshow(im);
title(char(labels(i)),'Color',colorText);
end
end
end
xlabel('ALL mismatch incidents')
%start plotting sample of correct match
figure('Name','Sample of Correct Match')
true=0;
false=0;
for i = 1:size(imdsValidation.Files,1)
ii = randi(size(imdsValidation.Files,1));
im = imread(imdsValidation.Files{ii});
if labels(ii) == imdsValidation.Labels(ii)
colorText = 'g';
true=true+1;
if true > f0*f1
29
break
else
subplot(f1,f0,true);
imshow(im);
title(char(labels(ii)),'Color',colorText);
end
end
end
xlabel('Sample of Correct Incidents')
%% Compute Accuracy
YValidation = imdsValidation.Labels;
accuracy02 = sum(labels ==
YValidation)/numel(YValidation);
display(accuracy02);
30
ACKNOLEDGEMENT
I wish to thank my parents for their support and encouragement throughout my studies.
31
REFERENCES
Online Sources
1. Online Article: Convolutional Neural Network Architecture: Forging Pathways to the
Future: https://missinglink.ai/guides/convolutional-neural-networks/convolutional-neural-
network-architecture-forging-pathways-future/
2. Dataset: IITD Iris Image Database
https://www4.comp.polyu.edu.hk/~csajaykr/IITD/Database_Iris.htm
