AGE ESTIMATION USING FACIAL IMAGES

HASHIM BOLAJI BELLO 1427571

Supervisor: Dr. Wahju Sediono

Co-supervisor: Asst Prof. Hazlina Md Yusof
Co-supervisor (2): Assist Prof. Noor Hazrin Hany
Mohamad Hanif

DEPARTMENT OF MECHATRONICS ENGINEERING

KULLIYYAH OF ENGINEERING

INTERNATIONAL ISLAMIC UNIVERSITY

MALAYSIA

MAY 2019

1
 ABSTRACT

Features of the human face holds a lot of data that can be processed into useful
information such as age, state of emotion, race and possibly what types of illness
different a human specimen suffers from. To extract all these information, proper
analysis of facial data is required through image processing of facial images. Due to
lack of availability of complete data on human face with their respective ages, it has
proved challenging to accurately determine the age of individuals based on just data
obtained from their faces.
In this paper, features of human face will be extracted using OpenCV library of
Python programming language, different functions and classes used to detect edges,
OTSU Thresholding and many other OpenCV functions. Specimen will consist of
both voluntary and online data obtained from data science website (kaggle.com),
facial images used in this paper are of people aged 18 years to 70 years. The obtained
data were used to train the deep convolution neural network algorithm to estimate the
age of facial images in a 4 years’ interval. Part of this research will be conducted
using a JavaScript framework called VueJs for the estimate age of each facial images
and compare them with result obtained using OpenCV and deep learning, JavaScript
will also be used for the design of the user interface of the resulting application. It
should however be noted that using JavaScript doesn’t require a lot of testing data.

ii
 TABLE OF CONTENTS

Abstract.........................................................................................................................ii
Table of Contents........................................................................................................iii
List of Figures..............................................................................................................iv
Approval Page...............................................................................................................v
Declaration...................................................................................................................vi
Acknowledgements.....................................................................................................vii

CHAPTER 1: INTRODUCTION...............................................................................1
1.1 Background......................................................................................................1
1.2 Statement of the Problem.................................................................................2
1.3 Objective of the Study.....................................................................................2
1.4 Scope of the Project.........................................................................................3
1.5 Significance of the Study.................................................................................4
1.6 Organization of the Report...............................................................................4

CHAPTER 2: LITERATURE REVIEW...................................................................5

2.1 Introduction......................................................................................................5
2.2 Related Work...................................................................................................6

CHAPTER 3 : METHODOLOGY.............................................................................8
3.1 Introduction......................................................................................................8
3.2 Face Detection.................................................................................................8
3.3 Face Alignment................................................................................................9
3.4 Face Age Feature Extraction..........................................................................11

CHAPTER 4 : RESULTS AND DISCUSSION.......................................................13

4.1 Results............................................................................................................13
4.2 Result and Discussion of Face Detection.......................................................13
4.3 Results and Discussion for Face Alignment..................................................15
4.4 Result for Face Age Feature Extraction and Age Estimation........................18

CHAPTER 5 : CONCLUSION.................................................................................20

REFERENCES...........................................................................................................21

iii
 LIST OF FIGURES

Figure 1. 1 Flowchart of Age estimation system (Ranjan et al., 2015) 9

Figure 1. 2 : Different face shapes used in shape based cues for age estimation 10Y
Figure 3. 1: Figure3: Face Detection system................................................................16
Figure 3. 2: (a) Trimmed image from detected face (b) Landmarks detected image (c)
Grayscale image...........................................................................................................17
Figure 3. 3 : Landmarks Boundaries..............................................................................1
Figure 4. 1: Face detection of a group of dancers........................................................21
Figure 4. 2: Grayscale images of detected faces..........................................................21
Figure 4. 3: Face Detection Programming Written in Python (OpenCV)....................22
Figure 4. 4: Original Images and Aligned Images........................................................23
Figure 4. 5: Figure 10: Face Alignment Programming Written in Python (OpenCV). 24
Figure 4. 6: Image showing different facial landmarks................................................25
Figure 4. 7: Image showing different facial landmarks................................................26

iv
 APPROVAL PAGE

I certify that I Have supervised and read this study and that in my opinion, it conforms

to acceptable standards of scholarly presentation and is fully adequate, in score and

quality, as a Final Year Project report as partial fulfilment for a degree of Bachelor of

Engineering (Mechatronics) (Honours).

……………………………………………………….

[DR. ING. WAHJU

SEDIONO]

supervisor

v
 DECLARATION

I hereby declare that this dissertation is the result of my own investigations,

except where otherwise stated. I also declare that it has not been previously or

concurrently submitted as a whole for any other degrees at IIUM or other institutions.

HASHIM BOLAJI BELLO

Signature: ……………………………. Date: ………………..

vi
 ACKNOWLEDGEMENTS

All praises are due to Allah for providing me with the strength and means to

carry out my first final year project. My great thanks and feelings of appreciation goes

to Dr. Wahju Sediono who gave me all the guidelines required to see me through this

research. Owning to his guidance and perseverance through the semester, I manage to

complete my research. I would also like to thank the coordinator of this program, Dr.

Noor Hazrin Hany Bt Mohamad Hanif and Dr. Hazlina Bt Md. Yusof for giving us the

right information at the right time.

vii
 CHAPTER 1

INTRODUCTION

1.1 BACKGROUND

The significance of facial age estimation is becoming increasing apparent, from

industrial and technical uses such as human computer interface to video surveillance,

for academic researches, the need for facial age estimation cannot be overstated.

Facial age estimation is the process of assigning an exact age or a range of ages to the

facial image of a human being based on cues and features that can be extracted on the

face (Ranjan et al., 2015).

Many different approaches have been used to estimate age, unfortunately they

are either inaccurate or largely resource intensive. In the medical field, laboratory

depend on the use of body fluid analysis such as blood test, cholesterol level, albumin

and other indicators to determine a person’s “physiological age” and determine the

level of aging in human being (Liao et al., 2018). However, medical method of

estimating human age is largely inefficient and very expensive equipment and

facilities are required.

In order to solve the problem of age estimation using image processing, an

algorithm is developed to extract features of the face by measuring the

anthropometrical features of different part of the faces is properly recorded, these

features include the nose, eyes and the lip. Equally important are shapes texture based

cues that gives information on the exact age or range of age of a facial image. The

algorithm works in 4 different steps as described by (Ranjan et al., 2015), face

1
detection, face alignment, feature detection and then trained based on facial images

with layers of neural network regression as shown in figure 1 below.

Figure 1. 1 Flowchart of Age estimation system (Ranjan et al., 2015)

1.2 STATEMENT OF THE PROBLEM

As mentioned earlier, the importance of age estimation technology is becoming more

apparent in our society. There are very limited available technologies that correctly

estimate the range of an individual’s age much less the exact age of the individual.

Inaccurate estimation of age are wrong data that would make decision made based on

them wildly inaccurate, these decisions include but not exclusively commercial

purpose (e.g. targeted advertisement), social media (e.g. tagging system on social

media) and medical purposes.

1.3 OBJECTIVE OF THE STUDY

1. To develop a system that extract all relevant data from image of a human face

2. To develop a modular system for the face change recognition caused by aging

process based on data extracted in 1.

3. To develop a system that estimate age based on live images.

2
 4. To evaluate the performance of the developed system

1.4 SCOPE OF THE PROJECT

This project will mainly cover exploring methods of estimating ages using a texture

and shape based cues that are obtained from facial images. Texture based cues include

the smoothness and roughness of the skin, the elasticity of the skin while shape based

cues are based on the symmetry of the face, the shape of the face, jaw area and

position of different organ (eyes, nose and mouth) of the face. In addition,

anthropometric measurement of the face (size of the mouth, length between the eyes,

the structure of the nose) will be extracted to give addition accuracy to the texture and

shape based cues. Furthermore, all the data extracted from the face using OpenCV

library will be used to train the deep learning convolution neural network. The trained

algorithm will be used to test fresh of set facial images to properly evaluate the

accuracy of the model or system

Figure 1. 2 : Different face shapes used in shape based cues for age estimation

3
1.5 SIGNIFICANCE OF THE STUDY

The importance of this project is to study the human face and all possible data that are

hidden in it. Study the setback of existing system and work on designing new system

with deep convolution neural network (DCNN) algorithms that corrects the deficiency

of existing age estimation methods and improves their efficiency.

1.6 ORGANIZATION OF THE REPORT

This report will include five chapters which include, The Introduction, Literature

Review, Research Methodology, Result and Discussion, and Conclusion.

Chapter 1 – Introduction: This chapter focuses on basic ideas of the research, the

problem statement, its objective and scope as well as the background of the project

Chapter 2 – Literature Review: The chapter goes in-depth to explaining meaning of

the research, giving detail information about the research and information obtained

from different sources largely academic sources

Chapter 3 – Research Methodology: The chapter will elaborate on the methods that

will be used in this research to achieve results and reach conclusions. It will include

measurement, calculations.

Chapter 4 – Result and Discussion: The chapter examines the result of experiences

and discuss them.

Chapter 5 – Conclusion: The chapter draws important conclusions from the research

and important recommendations.

4
 CHAPTER 2

LITERATURE REVIEW

2.1 INTRODUCTION

Automatic age estimation, which involves evaluating a person’s exact age or age

group is a very important topic in human face image comprehension. The role of

estimating exact human age adopts a close representation of the age labels, and the

task of age group estimation divides the label into a group. The research focuses on

determining the range of age for particle image or live video.

According to (Chang and Chen, 2015), to train an age estimating algorithm,

most approaches formulate it as either a multi-class classification problem or a

regression problem. Multi-class approaches simply treat the age values as independent

labels and train a classifier to infer the person’s age.

Instead of classification or regression, recent studies have proposed learning-

to-rank approaches to solve the age estimation problem recently. Unlike the

classification approaches that treat the labels naively as independent tags or the

regression approaches that treat the labels as simply proportional quantities, ranking

approaches can be trained by adopting the ordering property of the labels, thereby

achieving superior performance (Chang and Chen, 2015).

In an effort to close the gap between the capabilities of age and gender

estimation systems and face recognition systems (let alone the gap between computer

and human capabilities) we take the following steps. First, we offer a public data-set

of labeled images, and associated benchmark protocols, designed to reflect the

5
challenges of real-world age and gender estimation tasks. Besides being massive in

the number of images and subjects it includes, it is unique in the nature of its images.

These were obtained from online image albums captured by smart-phones and

automatically uploaded to Flicker before being manually filtered or “curated”. It

therefore contains many images which would typically be removed by their owners as

being faulty or otherwise unsuitable (Eidinger, Enbar and Hassner, 2014).

Eidinger, Enbar and Hassner (2014) postulated an alternative approach that

seeks to model the high dimensional region in some feature space, which characterizes

how facial appearances change with age. To this end, previous methods have used

subspaces [19] or high dimensional manifolds [14]. Similar to the anthropometric

models, however, they assume accurate alignment of faces, and the underlying

assumption of the manifold structure of facial appearances across ages, which may not

strictly be true.

2.2 RELATED WORK

Early work on age estimation, inspired by studies of aging in biology, proposed

representing faces by obtaining facial measurements that change with a person’s age.

These so-called anthropometric models, consider measurements taken from different

facial locations, and analyzed using expert tailored knowledge, represented as

manually defined statistical models, to determine age. More recently, [18] presented a

similar approach for modeling the age of young people. These methods, based on their

need for accurate localization of facial feature detections, may not be suitable for the

realistic scenarios considered here (Eidinger, Enbar and Hassner, 2014).

Convolutional layers with nonlinear activation functions compose CNN. The output

convolutions are computed from a previous input layer. Convolutional layers apply

6
distinct filters and associate their results with the output. More specifically, the layer’s

parameters consist of a set of learnable filters, operating on small receptive fields that

extend through the full size of the input. A sequence of these layers is responsible for

extracting features, and turning abstract low level features into high-level ones. The

network is responsible for learning the filter weights, allowing problems with different

image databases to use the same network. Nevertheless, the use of multiple

convolution layers will decrease the network processing speed, due to the number of

operations that each layer performs.

7
 CHAPTER 3

METHODOLOGY

3.1 INTRODUCTION

In this chapter, the methods and procedures for testing and experimentation applied

throughout the project would be explained in great detail. The procedures are

subdivided into 3 namely:

1. Face Detection

2. Face Alignment

3. Face Age Features extraction and Age Estimation

In this section, important areas of interest are discussed at great length.

3.2 FACE DETECTION

There are different types of Image dataset that can be obtained, a labelled dataset that

comes with the year each facial images are taken or an unlabeled dataset with no

information about when the facial image was taken, the former can be obtained from

online resources like flickr.com while the later can be obtained from data science

website (like kaggle.com). It is however important to note that regardless of the type

of image dataset used (labelled and unlabeled), the process for face detection is similar

(Hu et al., 2016). For age estimation, the first level of computation is to identify the

faces in every input image using OpenCV algorithm for face detection. Two main

modules are used for face detection, first an input image is passed into a pyramid of 7

levels, the resulting image pyramid is used by the network to generate a new pyramid

of 256 feature maps at the fifth convolution layer (conv5). A 3 X 3 max filter is

8
applied to the feature pyramid at a stride of one to obtain the max 5 layer which

integrates the fifth convolution layer, each face contains a bias which needs to be

removed, it is removed by normalizing the activations each level of the feature

pyramid, the resulting feature pyramid is called “norm5” (Ranjan et al., 2015). The

second module is trained on norm5 using a linear SVM. The root filter of a certain

width and height is convolved with the feature pyramid at each resolution level.

Scores are assigned to different region of images, scores above a certain threshold are

subjected to greedy non-maximum suppression to trim low scoring regions with

intersection overlap union (IOU) above 0.3 with a bounding rectangle to denote the

face region (Ranjan et al., 2015).

Figure 3. 1: Figure3: Face Detection system

Images contained in a facial image dataset often contain images of different sizes and

resolution, the above method of face detection is very advantageous because it can

detect images no matter how big or small the size and resolutions are.

9
3.3 FACE ALIGNMENT

Facial landmarks are different parts of the face that are identifiably unique to a face

that makes its detection possible, facial landmarks include but are not limited to right

eyes center, left eye center, nose base, the lip etc. Considering the detected face

bounded by the rectangle in the previous section. Face alignment algorithm first trims

the facial image and assigns an initial mean shape S0. The shape estimate S(t+1) at

(t+1)th can be written in terms of the shape estimate at the previous stage S t and a

shape regressor rt which determines the increment of the shape estimate using the

current shape estimate and the difference in pixel intensity values as the features

(Ranjan et al., 2015).

The equation below the regression process:

S(t+1) = S(t) + rt(I, S(t))

Once the facial landmarks have been detected as shown in figure 4 below, the facial

image is aligned into the canonical coordinate system with the similarity transform

using facial landmarks, the image is then converted into grayscale with a resolution of

256X256.

Figure 3. 2: (a) Trimmed image from detected

10 face (b) Landmarks detected image
(c) Grayscale image
Furthermore, different facial landmarks can be bounded by rectangle to individually

extract data with greater accuracy as shown below in figure 5.

Figure 3. 3 : Landmarks Boundaries

3.4 FACE AGE FEATURE EXTRACTION

Deep Convolution Neural Network has proven very efficient in extraction of facial

features from facial images and detected landmarks. Hence features extraction can be

computed using the CNN learnt from the face detected which give a vector

representation. The representation is tightly packed and discriminative for face. The

network is trained for a classification task on KAGGLE-WebFace Dataset with over

50,000 subjects as the classes. The aligned grayscale images obtained from the face

alignment procedure with resolution (shape) of 256X256 are used as input. The

DCNN used in this research consist of 10 convolution layers, 5 pooling layers and 1

fully connected layer. To estimate the age, 3 layer of neural network is used to learn

the regression function below.

11
Each layer is followed by a parametric linear rectified linear unit (PReLU) activation

function except the last layer that evaluates the age of the input facial image. The first

layer is the input layer which takes the 320 dimensional feature vector obtained from

the face-identification task. The output of this layer, after the dropout and PReLU

operation, is fed to the fist hidden layer containing 160 number of hidden units.

Subsequently, the output propagates to the second hidden layer containing 32 hidden

units. The output from this layer is used to generate a scalar value that would describe

the apparent age (Ranjan et al., 2015).

12
13
 CHAPTER 4

RESULTS AND DISCUSSION

4.1 RESULTS

The initial result obtained after testing the DCNN algorithm with different images are

explained in a modular fashion. First the result and possible discussion on the face

detection are enumerated, secondly the face alignment algorithm is discussed at length

and all the precaution that need to be taken for future studies of this nature. Lastly in

the section how the DCNN algorithm works to extract and estimate the age of each

image specimen is discussed.

4.2 RESULT AND DISCUSSION OF FACE DETECTION

The face recognition aspect is largely done using OpenCV Python library. The

“detectMultipleScaleFunction” method used in the code in the figure below detects

the face image which is in grayscale mode, the scale factor is used to adjust for some

images that may be too close to the camera and as a result will suffer from poor

exposure which may lead to images not be been detected or non-facial images

detected as a facial image. The number of faces detected by the algorithm is determine

by the “minSize” method.

14
 Figure 4. 1: Face detection of a group of dancers

In this instant of testing, the algorithm detected a non-facial part of the body as a

facial image, this could be due to multiple issues, first is the quality of the camera, it

produces a high resolution image. The picture was taken far away and in area of high

light exposure that interferes with lens of the camera. All these can be corrected by

adjusting the parameter in the “scaleFactor” method in OpenCV. Subsequent testing

that were carried out on the algorithm were accurate to a significant degree.

Figure 4. 2: Grayscale images of detected faces

15
 Figure 4. 3: Face Detection Programming Written in Python (OpenCV)

4.3 RESULTS AND DISCUSSION FOR FACE ALIGNMENT

The importance of alignment has been properly stated in the methodology section.

Given a set of image data, the matrix shape of each individual image is different, the

size and the color map (e.g. BGR, RGB, GRAY) for each individual image is

different, Alignment is the process of making all images in a dataset into one unique

form from which proper age estimation computation can be carried out. Without face

alignment, one would have to write different algorithm for different faces based on

their unique individual properties (size, shape, orientation and color map).

16
To successfully align every images in the dataset, the image must be centered, rotated

such that the landmarks such as the left eye center and the right eye center are

horizontally parallel, the nose base and the lip are also vertically parallel to each other.

Below are results of some alignment testing that was carried out on random images

form the internet.

Figure 4. 4: Original Images and Aligned Images

It is worth noting that some images in the dataset may be properly the aligned and

there will be no need to re align them. In cases like this, it is important to save

computing resources by not computing them in the face alignment algorithm. Hence a

17
program is written to check if each images in the dataset are aligned and if they are,

they are not computed for face alignment.

Figure 4. 5: Figure 10: Face Alignment Programming Written in Python (OpenCV).

18
4.4 RESULT FOR FACE AGE FEATURE EXTRACTION AND AGE
ESTIMATION

Using CNN, different factors that can affect the appearance of a human being are

taken into consideration for example, people in different part of the world age

differently based on the weather condition of their origin, skin color and so on.

Moreover, emotion plays a huge role in the way people appear, as a result extracting

features requires proper detection of landmarks on the face as shown in the figure 11

below.

Figure 4. 6: Image showing different

facial landmarks

The estimated ages were in range of 4-5 years because the difference in face attributes

over a period of one year is very negligible, hence, the study will be more productive

19
with longer age ranges. Other important factors noted in age extraction are monitoring

the activity of the face and change in emotions of the face and the engagement of the

face with the camera.

Figure 4. 7: Image showing different facial landmarks

Finally, the UI of the system was built with a JavaScript frameworks called Vuejs. It

makes it possible to host the whole system on a web browser since JavaScript is a web

language. The importance of hosting it on a website is to make it more accessible,

20
which allows more image data to be tested on the system thereby increasing its

efficiency.

Chapter 5

CONCLUSION

This is the first part of the studies, the objective highlighted in the introduction are all

achieved except monitoring aging process. The scope covered in this studies is a

precursor to monitoring ageing process in different human being. Monitoring aging

process requires a lot of time and resources particularly computing power and

resources because it requires a lot of steps which including classifying all images

based on ethnicity, skin color, texture and shape of the face. Since all this classes

respond differently to aging process, different learning algorithm has to be developed

for them.

The result obtained in this experiment are based on voluntary testing of about 100

IIUM students and 10,000 facial images. Due to the low level of data sample space the

accuracy of the initial system is still at a minimum. For future studies, higher level of

data will be used to conduct the research. DCNN algorithm only get better when more

data is analyzed.

Lastly, the proposition made in this studies should be recommended to be more

deeply researched and tested for further possible improve and possible implementation

in every facet of life.

21
 REFERENCES

Chang, K. and Chen, C. (2015) ‘A Learning Framework for Age Rank Estimation
based on Face Images with Scattering Transform’, IEEE TRANSACTIONS ON
IMAGE PROCESSING, 7149(c), pp. 1–14. doi: 10.1109/TIP.2014.2387379.
Eidinger, E., Enbar, R. and Hassner, T. (2014) ‘Age and Gender Estimation of
Unfiltered Faces’, IEEE TRANSACTIONS ON INFORMATION FORENSICS
AND SECURITY, 9(12), pp. 2170–2179.
Hu, Z. et al. (2016) ‘Facial Age Estimation with Age Difference’, IEEE
TRANSACTIONS ON IMAGE PROCESSING, 7149(c), pp. 1–11. doi:
10.1109/TIP.2016.2633868.
Liao, H. et al. (2018) ‘Age Estimation of Face Images Based on CNN and Divide-
and-Rule Strategy’, Mathematical Problems in Engineering, 2018, pp. 1–8.
Ranjan, R. et al. (2015) ‘Unconstrained Age Estimation with Deep Convolutional
Neural Networks’, IEEE International Conference on Computer Vision
Workshops, pp. 109–117. doi: 10.1109/ICCVW.2015.54.

22