dinh quang huy thesis
TRANSCRIPT
Automatic Discovery of ConnectionBetween Vietnamese’s Anthropometric
Features
Dinh Quang Huy
Faculty of Information TechnologyCollege of Technology
Vietnam National University, Hanoi
Supervised byAssociate Professor Bui The Duy
A thesis submitted in fulfillment of the requirements for the degree of
Master of Information Technology
December, 2010
2
ORIGINALITY STATEMENT
‘I hereby declare that this submission is my own work and to the best of my
knowledge it contains no materials previously published or written by another
person, or substantial proportions of material which have been accepted for the
award of any other degree or diploma at University of Engineering and Technology
or any other educational institution, except where due acknowledgement is made
in the thesis.
I also declare that the intellectual content of this thesis is the product of my own
work, except to the extent that assistance from others in the project’s design and
conception or in style, presentation and linguistic expression is acknowledged.’
Signed ........................................................................
i
Abstract
Long time ago, when people found a skeleton, it was hard to determine
who the victim was. However, people are trying to find a way to solve this
problem due to its demands and importance. Several methods have been
introduced for identifying deceased persons, some more effective than others.
Facial reconstruction is one of them. It is a work of recreating the face of
a person from his skeletal remains. At the first days, facial reconstruction
is done using clay, where a skillful experts who understand the structure of
skull and skin very well to use clay to build up the depth of tissue on the
skull to that of a living individual. Later, this method is computerized and
people tend to develop 3D facial reconstruction systems.
In the facial reconstruction systems, the most important issue is to predict
the soft tissue depths at every location or some locations. Most researches
try to obtain a database of soft tissue thicknesses at facial landmarks, and
store the average thickness for every landmark. When performing the re-
construction, these thicknesses are referenced, and the face is built based on
the skull model. Their approaches have some problems in data collecting,
and they do not make use of the discovered skull to predict the thicknesses.
Therefore, the accuracy is very low and most of the time, they need to man-
ually modify the model generated from the system a lot in order to receive
a suitable face.
Realizing that the soft tissue thickness and some other anthropometric
features may have some relationships with the skull shape, we propose a
method for automatic discovery of these connections. We first collect data
using the CT technique which is the most accurate method at the moment.
After that, we try some machine learning techniques on the data to see the
performance. The evaluations and comparison with other approaches are
also given in the thesis.
ii
Table of Contents
1 Introduction 1
1.1 Overview and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Our Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Background 4
2.1 Previous Work in Facial Reconstruction From Skulls . . . . . . . . . . 4
2.1.1 2D Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.2 Manual 3D Reconstruction . . . . . . . . . . . . . . . . . . . . 6
2.1.3 Computer-Aided Reconstruction . . . . . . . . . . . . . . . . . 7
2.2 Facial Reconstruction Systems . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 System developed by Bjorn Anderson, Martin Valfridsson in
2005 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.2 System developed by Kolja Kahler and Jorg Haber . . . . . . 9
2.2.3 FACES - software developed by Salerno University, Italy . . . 10
2.3 Facial Landmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Important Facial Features . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4.1 Ears . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.2 Eyes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.3 Nose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.4 Lips . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Soft tissue thickness studies . . . . . . . . . . . . . . . . . . . . . . . 13
2.6 Available Soft Tissue Thickness Data . . . . . . . . . . . . . . . . . . 15
3 Automatic discovery of connections between Vietnamese’s anthro-
pometric features 16
3.1 Data description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Data collecting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
iii
iv TABLE OF CONTENTS
3.3 Discovery of anthropometric relationships using linear regression . . . 23
3.4 Discovery of anthropometric relationships using neural networks . . . 25
3.4.1 Select network structure . . . . . . . . . . . . . . . . . . . . . 25
3.4.2 Initialize and train the network . . . . . . . . . . . . . . . . . 26
4 Evaluation and Result 29
5 Conclusions and Future Work 35
List of Figures
2.1 Matching skull into drawing portrait . . . . . . . . . . . . . . . . . . 5
2.2 Matching skull into a picture . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Successful clay reconstrion by LSU Faces Lab . . . . . . . . . . . . . 6
2.4 Process of Reconstruction using volumetric data . . . . . . . . . . . . 8
2.5 Result of Bjorn Anderson and Martin Valfridsson’s reconstruction . . 9
2.6 Facial Reconstruction Diagram by FACES . . . . . . . . . . . . . . . 11
2.7 Facial landmarks Location . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1 Phillip MX8000D CT Scanner . . . . . . . . . . . . . . . . . . . . . . 19
3.2 CT images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Head CT image taken with sagittal plane . . . . . . . . . . . . . . . . 20
3.4 Head CT image taken with verticle plane that goes through the mid-
dle of the left eye socket . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.5 Head CT image taken with vertical plane that goes through the forehead 22
3.6 Head CT image taken with horizontal plane that goes through the
zygion landmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.7 Head CT image taken with horizontal plane that goes through the
gonion landmarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.8 Example of linear regression . . . . . . . . . . . . . . . . . . . . . . . 24
3.9 A feed-forward network with a single output layer (a) and with one
hidden layer and one output layer (b) . . . . . . . . . . . . . . . . . . 26
3.10 A recurrent network with hidden neurons . . . . . . . . . . . . . . . . 26
3.11 Neural network structure used in the study . . . . . . . . . . . . . . . 27
4.1 Regression results obtained by ten-fold cross validation for pronasale
thickness using (a) neural network model and (b) linear regression
model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
v
vi LIST OF FIGURES
4.2 Regression results obtained by ten-fold cross validation for nose length
using (a) neural network model and (b) linear regression model. . . . 31
4.5 Regression results obtained by ten-fold cross validation for upper lip
border using (a) neural network model and (b) linear regression model. 31
4.3 Regression results obtained by ten-fold cross validation for nose height
using (a) neural network model and (b) linear regression model. . . . 32
4.4 Regression results obtained by ten-fold cross validation for pupil-
pupil distance using (a) neural network model and (b) linear re-
gression model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.6 Regression results obtained by ten-fold cross validation for lower lip
border using (a) neural network model and (b) linear regression model. 32
4.7 Facial Reconstruction Result Using Linear Regression Equations . . . 33
4.8 Matching the face and the skull . . . . . . . . . . . . . . . . . . . . . 33
List of Tables
2.1 List of Facial Landmarks . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1 Input Data Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Target Data Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.1 MSE values for ’average method’ (AVG), Linear Regression (LR), and
Neural Network (NN). The best performance is in boldface. . . . . . . 30
4.2 Equations for linear correlation between input and output, with the
corresponding MSE when applied with the whole data set. In the
equations, x is the input and y is the output. . . . . . . . . . . . . . . 34
vii
Abbreviations
3D Three-dimensional
CT Computed Tomography
MRI Magnetic Resonance Imaging
2D Two-dimensional
RBF Radial Basis Functions
MSE Mean Square Error
viii
Chapter 1
Introduction
1.1 Overview and Motivation
Facial reconstruction is the work of recreating the face of an individual from his dis-
covered skull. This process is mainly used in criminal investigations to facilitate vic-
tim identification when no other means are available. Besides, facial reconstruction
is also used in archaeology to verify the remains of historic figures or in anthropology
to approximate the look of prehistoric hominids.
People have been recreating the face of an unidentified individual from their
discovered skulls for nearly a hundred years. At the first days, facial reconstruction is
done using clay. This method requires skillful experts who understand the structure
of skull and skin very well to use clay to build up the depth of tissue on the skull
to that of a living individual. The experts first place the landmark dowels on the
pre-defined craniofacial landmarks on the skull. After that, clay is applied and the
expert interpolates with clay between the landmark dowels to build up the skin. This
method is called the Krogman method [Kro46] and is still used in non-automatic
forensic facial reconstruction now. The expert skill and amount of time required
have motivated researchers to try to computerize the technique. A well-designed
computer-aided facial reconstruction system has many advantages, including great
reduction in time consumption. Using such a system, we can produce several possible
facial models from a given skull by using parameters determining the person’s age,
weight, and gender.
Recently, the rapid development of 3D equipments and technology enable us to
advance into this field of research. A lot of computerized methods for 3D facial
1
2 Chapter 1. Introduction
reconstruction have been proposed and developed, which make use of computer
program to transform 3D scanned models of the skull into faces. Many researches
try to follow the manual approach, but make use of the computer to fasten the
process of reconstruction. In these methods, they have the discovered skull scanned
to get its 3D model. After that, they need to calculate the soft tissue thicknesses at
every location at the skull surface to build the final facial model and the most critical
issue turns out to be discovering these thicknesses. In order to solve this problem,
researchers tend to collect a database of information about skin-skull models and
then make use of it, but differ in the way of collecting and processing. Mostly,
they average the thicknesses for every record in the database and use these for
every discovered skull. The preview of these related works are described in the
next chapter. The simplicity of the methods using for calculating these thicknesses
leads to low performance in the facial reconstruction systems. In addition, some
important facial features such as nose and eye shapes are often reconstructed without
any information from the skull shape. Realizing this drawback in nowadays facial
reconstruction systems, we aim to propose a method for better prediction of the
facial information. This study is made based on Vietnamese data, but can be applied
to any other race with proper data collection.
1.2 Our Contributions
In scope of a Master thesis, we propose a method for automatic discovery of connec-
tions between anthropometric features such as tissue thicknesses, distance between
two pupils, nose height and the skull shape. This work includes the method for
database collecting, and how we apply machine learning to discover the relation-
ships. The connection model then can be used to solve the problem of reconstruct-
ing the face from the skull. Our results from evaluations also prove that there is
relationships between facial soft tissue thickness and the skull shape.
1.3 Thesis Organization
The rest of this thesis is organized as follows. Chapter 2 presents some background
knowledge that is involved. The issues in this chapter include of previous work in
facial reconstruction, current facial reconstruction systems. We also review the work
1.3. Thesis Organization 3
of soft tissue thickness studies and available soft tissue thickness data. Chapter
3 describes the proposed method of automatic discovery of connections between
Vietnameses anthropometric features. Our method of data organization, how to
collect this data, and how we make use of this data are provided in this chapter.
Chapter 4 presents our results and the comparison with other approaches. Chapter
5 concludes our work and gives the future research directions based on the results
obtained so far.
Chapter 2
Background
2.1 Previous Work in Facial Reconstruction From
Skulls
Reconstructing a face from a skull is to predict a face as much accurately as possible
from the discovered skull. There are many different approaches in order to solve
this problem. These approaches can be divided into three main techniques, 2D
reconstruction, manual 3D reconstruction, and 3D reconstruction using computers.
2.1.1 2D Reconstruction
This technique requires a forensic artist who draws the picture of the reconstructing
face. There are two main 2D approaches, matching the skull into the drawing
portrait and matching the skull into pictures or videos. In the case of matching the
skull into the drawing portrait, the artist first draws a simple version portrait based
on the skull’s metrics. After that, he matches this drawing into the skull image and
watch. He then redraws or fixes some features until the skull image and the drawing
fit perfectly. The process of this technique is shown in Figure 2.1.
The other technique, matching the skull into a picture or video, is used when
people want to compare the face with the skull to identify the correlations or if the
skull belongs to the people in the picture or video (Figure 2.2).
4
2.1. Previous Work in Facial Reconstruction From Skulls 5
Figure 2.1: Matching skull into drawing portrait
Figure 2.2: Matching skull into a picture
6 Chapter 2. Background
Figure 2.3: Successful clay reconstrion by LSU Faces Lab
2.1.2 Manual 3D Reconstruction
The most common manual approach to facial reconstruction is the clay modeling
approach. In this method, people first put the landmark dowels on the predefined
craniofacial landmarks on the skull. The lengths of these dowels are defined using
one of the available soft tissue thickness data in the beginning. After that, they
attach clay to the skull in a way that the clay covers the dowels but still keeps a
continuous surface. This method is also known as the Krogman [Kro46] method,
and it gives good result without any help of a computer. However, this method
requires an expert with very good skill and still it takes a lot of time for a version
of reconstruction. Faces Laboratory in Louisiana State University, USA has some
successful examples of clay reconstruction as shown in Figure 2.3.
2.1. Previous Work in Facial Reconstruction From Skulls 7
2.1.3 Computer-Aided Reconstruction
Many researchers have worked on the facial reconstruction problem and have pro-
vided different solutions using computers. Mark Jones [Jon01] uses volumetric data
and cross correlation techniques. Matthew Cairns [Cai00] uses statistical tools such
as Generalized Procrustes Analysis and Principal Components Analysis. Katrina
Archer [Arc97] tries to computerize the manual facial reconstruction process. An-
other approach is presented by Kahler [KHS03] in which he use a physics based head
model with skin surfaces, virtual muscles, a mass spring system and landmarks to
reconstruct the face. Bullock [Bul99] uses the Krogman method for placing virtual
dowels on the digitized skull with the emission-based implicit modeling. In this mod-
eling, every polygon of the skull model emits a real value which is the interpolated
soft tissue thicknesses at the landmarks associated with the polygon.
There are many other techniques that share the same method of following steps.
The first step is generating the 3D model of the skull. We can do that with the
help of digitalized equipments such as the CT scanner or MRI scanner. The CT
technique enables accurate measurement of facial soft tissue thickness and is used
widely in collecting soft tissue thickness data. In addition, we can generate the 3D
model of the skull from CT images conveniently. The second step is to locate the
landmarks at the skull surface and the tissue depths at the landmarks. Based on
these landmarks, we can apply some regression techniques such as the RBF, B-spline,
and Non-uniform rational B-spline to generate the 3D grid of the skin surface. The
last step is to refine the reconstructed face, by tuning the features such as the eyes,
nose, ears and lips. This is a hard work because these features are unpredictable by
the skull shape only.
Figure 2.4 shows the process of the facial reconstruction proposed by Mark
Jones [Jon01] in which he compares volumetric data of the 18 remains with that
of a reference head. Firstly, the discovered skull is scanned using a CT scanner to
obtain volumetric data. After that, a reference head having the same sex, racial
and age characteristics as the discovered skull is chosen. Then a correspondence is
created between the two heads using correlation techniques. Finally, using this cor-
respondence, the soft tissue from the reference head is mapped onto the discovered
skull to produce the face of the unknown person.
8 Chapter 2. Background
Figure 2.4: Process of Reconstruction using volumetric data
2.2 Facial Reconstruction Systems
2.2.1 System developed by Bjorn Anderson, Martin Val-
fridsson in 2005
This system uses the 3ds max software and contains 9 steps. The first step is to
collect data from CT. In this step, the skull is scanned in a CT scanner and the CT
slices are stored in DICOM format. The second step is data segmentation, in which
CT slices are edited using segmentation software to remove artifacts such as metal
cavity fillings. This segmentation software is also used to produce a 3D model to
import into 3ds max. In the third step, the model is imported into 3ds max and
some pre-processing such as normalization and rotation is performed. In the forth
2.2. Facial Reconstruction Systems 9
Figure 2.5: Result of Bjorn Anderson and Martin Valfridsson’s reconstruction
step, the landmark dowels are located by users using the GUI. In the fifth step,
the holes in the cranium are covered. Step 6 is to perform mesh calculations. In
this step, based on the tissue depths at the landmark, the tissue depths at other
locations are calculated. The chin and neck are also constructed at this step. Step 7
is the creation of nose, eyes, ears and lips. Step 8 is the post process, in which they
alter the model based on knowledge about human faces. The last step is to insert
texture and render the final images. The sample result of this system is shown in
Figure 2.5.
2.2.2 System developed by Kolja Kahler and Jorg Haber
In December 2003, Kolja Kahler defended his PhD thesis with the title of ”A Head
Model with Anatomical Structure for Facial Modeling and Animation”. In the
10 Chapter 2. Background
thesis, he studied the facial muscles, how they work and deform, and built a facial
reconstruction software. This software is then developed [KHS03] by him and his
colleagues at the University of Saarland to become a complete system. Beside fast
reconstruction, the system also enables changing the emotion of the face, based on
the 24 types of muscle changes.
2.2.3 FACES - software developed by Salerno University,
Italy
This software performs facial reconstruction from the skull of prehistoric hominids.
The reconstruction is based on warping and deforming the template face chosen from
a set of models with different sexes, races. Some warping algorithms are applied to
carry out this work. The software uses two databases, craniometrical database of
skulls and pictorial physiognomic database of faces. Figure 2.6 shows the software’s
diagram.
2.3 Facial Landmarks
As described, most method uses the pre-defined points on the head which are called
facial landmarks. There are a number of different landmark configurations with
various numbers of landmarks. The mostly used landmarks are the 32 landmarks
described by Rhine [Rhi84]. The list of these landmarks and its position are de-
scribed in Table 2.1 and Figure 2.7. In our research, we make use of some extra
landmarks such as exocanthion, endocanthion, alare, pronasale, basion, subnasale,
and stomion. There are two types of landmarks, craniofacial and cephalometric.
They are the marks on the skull and corresponding marks on the skin surface,
respectively. In most data measurements, the distances between these pairs are
measured and stored.
2.3. Facial Landmarks 11
Figure 2.6: Facial Reconstruction Diagram by FACES
Figure 2.7: Facial landmarks Location
12 Chapter 2. Background
Table 2.1: List of Facial Landmarks
Number Landmark Name
1 Supraglabella
2 Glabella
3 Nasion
4 Rhinion
5 Mid Philtrum
6 Upper Lip Margin
7 Lower Lip Margin
8 Chin-Lip Fold
9 Mental Eminence
10 Beneath Chin
11 Frontal Eminence
12 Supraorbital
13 Suborbital
14 Inferior Malar
15 Lateral Orbit
16 Zygomatic Arch, midway
17 Supraglenoid
18 Gonion
19 Supra M2
20 Occlusal Line
21 Sub M2
2.4 Important Facial Features
Facial reconstruction’s aim is to produce a model of an individual’s face that can
be recognized by close friends or family members. Research presented by Wilkin-
son [Wil04] shows that hair, face outline, eyes, nose and mouth are the important
features for the recognition of faces.
2.5. Soft tissue thickness studies 13
2.4.1 Ears
Wilkinson [Wil04] stated that we cannot estimate the shape of the ears including
their size, form, and projection based on the skull only, because there are no un-
derlying bone that describes the ear appearance. However, the ear shape is not an
important feature in our face.
2.4.2 Eyes
There are some researches which show the eyes relationship. Firstly, for the depth
placement of the eyeball in the socket, Wilkinson [Wil04] indicates that a straight
line between the superior an d inferior orbital margins will touch the front of the
cornea. He also states that the opening of the eye is 60 to 80 percent of the width
of the orbit.
2.4.3 Nose
We can reconstruct the nose by looking at the shape of the nasal bones and cartilage.
Wilkinson [Wil04] shows a way to calculate the shape of the nose based on the angle
of the nasal bone. However, the cartilage part of the nose is hard to predict.
2.4.4 Lips
This is an important feature and it is said that the lips are determined by the
structure of the underlying bones. For example, a person with big teeth is likely to
have thick lips, while a person with small teeth often has thin lips.
2.5 Soft tissue thickness studies
As mentioned, the soft tissue thickness data play an important role in facial recon-
struction, whatever our method is. The issue of collecting soft tissue thickness data
to clarify the relationship between soft tissue and the underlying bony structure of
skull has been discussed by forensic medicine experts for more than a hundred years.
In 1883, Welcker [Wel83] obtained a database of soft tissue thicknesses by inserting a
thin blade into facial skin of cadavers at selected anatomical landmarks. After that,
he measured the depth of the blades penetration. Until middle 1980, all studies that
14 Chapter 2. Background
need to collect soft tissue thicknesses data at anatomical landmarks used cadaverous
populations and this ’needle technique’. However, this type of approaches has some
problems. First of all, a dead person’s tissues are not the same as in life due to
drying and embalming. Secondly, the skin can be deformed due to the penetration
of the needle. Lastly, it is hard to find the landmarks correctly through soft tissue
when performing the needle insertion. Since we need to produce a model as accurate
as possible, all these matters must be taken into consideration.
The ’needle technique’ cannot be used on living subjects, which leads to er-
rors in measurement. After 1980, with the development of technology, non-invasive
medical systems become popular. A variety of methods have been used to mea-
sure tissue depth in living subjects, including ultrasound, MRI, and CT. In 1987,
George [Geo87] used lateral craniographs to record the depths of tissue at the midline
anthropometric points. And in 2000, Manhein et. al. [MLB+00] used ultrasound to
collect information for sample of children and adults of sexes, varying ages and differ-
ent varieties. El-Mehallawi and Soliman [EMS01] and De Greef et. al. [DGPV+06]
also used ultrasound to conduct study. In 2002, Sahni et. al. [SJG+02] used MRI
to obtain tissue depth data of Indians. The most accurate measurement can be
obtained by using CT. This technique is faster and more accurate as it gives high
quality images. With the help of the computer, we can also construct the 3D model
from the CT images. In 1996, Phillips and Smuts [PS96] used CT technique to
obtain data of mixed population of South Africa. There are many more related re-
searches that collect soft tissue thicknesses for study. However, most measurements
are collected from rather small populations due to the harm it may cause when tests
are carried out. Ultrasound techniques seem to be the most accurate and safe as it
can be used without any considerable threat for the candidate [Wil04].MRI has the
advantage of collecting data in 3D format. Soft tissue visualization is excellent, but
bony tissue is not as well visualized as on the CT scan [VPST+07]. In addition, they
just gather tissue depths data at anthropometric landmarks, but give no informa-
tion about any relationship between these depths and the skull shape. Therefore, in
most facial reconstruction systems, they just use the average thicknesses that can
be calculated from the database for every landmark.
There are some researches which are available for Vietnamese. These researches
are made by Le Viet Vung (2005), Xu Xuan Khoi (1996), Le Gia Vinh (2005), Pham
Huu Phung and Nguyen Trong Toan (2007) and are provided in form of average
values and variations. They have also made some conclusions about Vietnamese
2.6. Available Soft Tissue Thickness Data 15
facial characteristics. For example, they conclude that Vietnamese faces are short
and wide type, Vietnamese noses are normal type. These researches are meaningful
when we want to verify a predicted model of a face or when we manually tuning the
facial features. However, this information is far from enough for an automatic facial
reconstruction system with accurate soft tissue thickness prediction.
2.6 Available Soft Tissue Thickness Data
There are several published soft tissue thickness data collections. Some datasets
of American Blacks and American Whites are provided by Rhine [Rhi84]. These
datasets are divided into groups of different sex and weight, and show the average
soft tissue thicknesses at Rhine’s landmark for each population in each group. Many
later facial reconstruction systems used these data collections as the method to define
the tissue depth. However, these datasets were obtained from cadavers that they
suffer the disadvantages described above.
In 2000, Manhein et. al. [MLB+00] published a study made on American Blacks
and Caucasian Americans using the ultrasound technique. This data is divided into
groups of age, with the landmarks similar to the ones made by Rhine. However,
the correctness of this data is considered higher than Rhine’s due to the method of
ultrasound in obtaining.
The latest dataset seems to be the T-tables (Tallied Facial Soft Tissue Depth
Data) provided by Stephan [SC10]. The T-tables represent pooled soft tissue depth
means from many of previously published studies. They started in 2008 and are well
updated up to now. The T-tables provide three sets of soft tissue thickness data
with different age range, 0 to 11 years, 12 to 17 years, and 18 years and beyond.
In contrast to any single soft tissue depth study that typically includes fewer than
40 individuals, each of the T-tables report values for more than 3000 individuals.
Therefore, the T-tables have an advantage that tolerates the measurement error in
single study. However, some researches [HLW85] [Dum86] have shown that race,
sex, age and weight have moderate impacts on soft tissue thickness. The T-tables
divide data into groups of age only, which make it hard to use. In addition, the data
provided is the already averaged one, so we cannot separate the data into different
groups.
Chapter 3
Automatic discovery of
connections between Vietnamese’s
anthropometric features
We treat the soft tissue thickness prediction issue the missing data problem, so that
the solution is straight forward. We need to build a database of sets of input and
target, with the input is the skull and target is the soft tissue thicknesses. After
the database is ready, the data processing begins. We try two approaches to find
the relationships: one using the simple linear regression, and one using the neural
network.
This chapter is organized as follow. Firstly, we describe our database such as
the features that we are storing. Secondly, we show how to collect this information
using the CT technique. Finally, we in turn present the data processing stage using
linear regression and neural networks.
3.1 Data description
Our database is the mean of storing information of candidates. For each candidate,
the information is divided into two sets, input set and target set. Since we cannot
store the entire skull model into database, only the ’important’ distances are stored.
These are the distances between landmark points on the skull and some skull metric
such as cranial height and cranial length. The target set is mostly the thicknesses at
landmark locations. In the case of facial reconstruction, only the input set is known.
16
3.1. Data description 17
Table 3.1: Input Data Fields
N# Name Description
1 cranial height distance from the midpoint of the anterior border of
the foramen magnum (basion) to the intersection of
the coronal and sagittal sutures (bregma).
2 n-rh distance between nasion and rhinion landmarks.
3 base nose length distance between the nasion and mid philtrum land-
marks.
4 cranial length distance from the midsagittal plane from the most
anterior point on the frontal (glabella) to the most
posterior point on the occipital (opisthocranion).
5 cranial breadth greatest width between the parietal eminences (eu-
ryon).
6 ex-ex distance between the left exocanthion landmark and
the right exocanthion landmark.
7 en-en distance between the left endocanthion landmark
and the right endocanthion landmark.
8 molar-molar distance between two molars.
9 al-al distance between the left alare landmark and the
right alare landmark.
10 nose socket width the largest horizontal distance of the nose socket.
11 forehead width distance between the left frontal eminence landmark
and the right frontal eminence landmark.
12 facial width (Zy-Zy) distance between the left Zygomatic Arch landmark
and the right Zygomatic Arch landmark.
13 jaw width (go-go) distance between the left gonion landmark and the
right gonion landmark.
14 upper face height (n-pr) distance between the nasion landmark and the
pronasale landmark.
15 n-ba distance between the nasion landmark and the basion
landmark.
16 ba-pr distance between the basion landmark and the
pronasale landmark.
17 facial height distance from the most anterior inferior point of the
mandible in the median sagittal plane (gnathion) to
the point of intersection of the internasal suture with
the nasofrontal suture (nasion).
18Chapter 3. Automatic discovery of connections between Vietnamese’s
anthropometric features
Table 3.2: Target Data Fields
N# Name Description
1 vertex soft tissue thickness at vertex landmark.
2 trichion soft tissue thickness at trichion landmark.
3 glabella soft tissue thickness at glabella landmark.
4 nasion soft tissue thickness at nasion landmark.
5 rhinion soft tissue thickness at rhinion landmark.
6 pronasale soft tissue thickness at craniofacial pronasale land-
mark.
7 nose length distance from nasion to cephalometric pronasale
landmark.
8 subnasale soft tissue thickness at subnasale landmark.
9 upper lip border soft tissue thickness at upper lip margin landmark.
10 lower lip border soft tissue thickness at lower lip margin landmark.
11 stomion soft tissue thickness at stomion landmark.
12 metal soft tissue thickness at metal landmark.
13 meton soft tissue thickness at meton landmark.
14 opisthooranion soft tissue thickness at opisthooranion landmark.
15 exocanthion (R) soft tissue thickness at right exocanthion landmark.
16 exocanthion (L) soft tissue thickness at left exocanthion landmark.
17 endocanition (R) soft tissue thickness at right endocanition landmark.
18 endocanition (L) soft tissue thickness at left endocanition landmark.
19 pupil-pupil distance between two pupils.
20 supraobital (R) soft tissue thickness at right supraobital landmark.
21 supraobital (L) soft tissue thickness at left supraobital landmark.
22 infraobital (R) soft tissue thickness at right infraobital landmark.
23 infraobital (L) soft tissue thickness at left infraobital landmark.
24 zygomatic arch (R) soft tissue thickness at right zygomatic arch land-
mark.
25 zygomatic arch (L) soft tissue thickness at left zygomatic arch landmark.
26 zygomatic (R) soft tissue thickness at right zygomatic landmark.
27 zygomatic (L) soft tissue thickness at left zygomatic landmark.
28 porion (R) soft tissue thickness at right porion landmark.
29 porion (L) soft tissue thickness at left porion landmark.
30 gonion (R) soft tissue thickness at right gonion landmark.
31 gonion (L) soft tissue thickness at left gonion landmark.
32 alare (R) soft tissue thickness at right alare landmark.
33 alare (L) soft tissue thickness at left alare landmark.
34 lateral nasal (R) soft tissue thickness at right lateral nasal landmark.
35 lateral nasal (L) soft tissue thickness at left lateral nasal landmark.
36 nose height projection of cephalometric pronasale landmark to
the skull surface.
37 bucal (R) soft tissue thickness at right bucal landmark.
38 bucal (L) soft tissue thickness at left bucal landmark.
3.2. Data collecting 19
The input fields are shown in Table 3.1 and the target fields are shown in Ta-
ble 3.2. Our data is collected from 220 candidates, 98 of which are males and 122
are females. The age range is from 17 to 82 and the weight range is from 38kg to
75kg.
3.2 Data collecting
The CT images which are used for our database is captured using a Phillip MX8000D
CT Scanner (Figure 3.1). This CT technique is very convenient because it is fast,
accurate, and can produce high quality images. It neglects the disadvantages of
[Kro46] because it can capture the skin and skull models of living objects. Besides,
we can build 3D models from CT images.
Figure 3.1: Phillip MX8000D CT Scanner
Figure 3.2 shows some samples of CT images that we are using. We have spe-
cialized software that we measure the distances visually by selecting the start point
and end point for each distance in the CT image. The image is at the correct ratio
with the real person so that the distance can be converted from pixel to milimetre.
20Chapter 3. Automatic discovery of connections between Vietnamese’s
anthropometric features
Figure 3.2: CT images
Firstly, we take the CT image for the sagittal plane (Figure 3.3) to measure the
soft tissue thickness for the landmarks: supraglabella, glabella, nasion, rhinion, mid
philtrum, upper lip margin, lower lip margin, chin-lip fold, mental eminence, and
beneath chin.
Figure 3.3: Head CT image taken with sagittal plane
After that, we take the CT image for the vertical plane that goes through the
middle of the left eye socket (Figure 3.4). From this image, we can measure the soft
3.2. Data collecting 21
tissue thicknesses for left frontal eminence, left supraorbital, and left inferior molar
landmarks. It works the same way with the right landmarks.
Figure 3.4: Head CT image taken with verticle plane that goes through the middleof the left eye socket
The CT image taken for the vertical plane that goes through the forehead (Fig-
ure 3.5) can be used to measure the soft tissue thicknesses at the left and right
porion landmarks.
The CT image taken for the horizontal plane that goes through the zygion land-
mark (Figure 3.6) is used to measure the soft tissue thicknesses at the left and
right zygion landmarks, and the CT image taken for the horizontal plane that goes
through the gonion landmark (Figure 3.7) is used to measure the soft tissue thick-
nesses at the left and right gonion landmarks.
22Chapter 3. Automatic discovery of connections between Vietnamese’s
anthropometric features
Figure 3.5: Head CT image taken with vertical plane that goes through the forehead
Figure 3.6: Head CT image taken with horizontal plane that goes through the zygionlandmarks
3.3. Discovery of anthropometric relationships using linear regression 23
Figure 3.7: Head CT image taken with horizontal plane that goes through the gonionlandmarks
3.3 Discovery of anthropometric relationships us-
ing linear regression
Linear regression is a method to model the relationship between two variables by
fitting a linear equation to the observed data. Linear regression is used popularly
in practical applications because models which depend linearly on their unknown
parameters are easier to fit than models which are non-linearly related to their
parameters. In prediction problems, linear regression is used to fit a predictive
model to an observed data set of y and X values. After that, this model can be used
to make a prediction of the value of y with any given value of X.
There are numerous methods to estimate the a and b parameters in the linear
equation such as the ordinary least squares, generalized least squares, least absolute
deviation, or maximum likelihood estimation. These methods differ in terms of
computational simplicity.
We use the simple linear regression which is the ordinary least squares estimator.
24Chapter 3. Automatic discovery of connections between Vietnamese’s
anthropometric features
Figure 3.8: Example of linear regression
A linear regression line has an equation of the form Y = a + bX, where X is one
of the distances in input set and Y is one of the thicknesses in output. The slope
of the line is b, and a is the intercept. Suppose there are n data points yi, xi. The
linear regression equation would give best fit for the data points in terms of least
square error. This error can be calculated using the following formular
lse =n∑
i=1
(yi − a− bxi)2 (3.1)
Since we consider the soft tissue thickness calculation the prediction problem,
we need to obtain such a linear equation for every output field. In order to deter-
mine which distance in the input data should be the X parameter, we try all the
possible fields in the input data. With each field, we apply the linear regression and
choose the one with best performance. This method can discover one-to-one linear
relationships effectively. Figure 3.8 shows an example of our process.
3.4. Discovery of anthropometric relationships using neural networks 25
3.4 Discovery of anthropometric relationships us-
ing neural networks
Artificial neural network have seen a rapid increase of interest over the last few years,
and are being successfully applied on a wide range of domains such as character and
speech recognition, signal processing. Neural networks have several advantages. The
most important advantage is the ability to learn from data and thus potential to
produce an acceptable output for previously unseen input data. Neural networks can
even work when input series contain low-quality or missing data. Another advantage
is the non linear nature. In addition, the network is very flexible to changes in the
environment. We only have to retrain the system in these cases.
The process of neural network design for prediction problems contains five pri-
mary steps, collect data, select network structure, initialize the weights and biases,
train the network, validate the network, and use the network.
3.4.1 Select network structure
The step of select network structure is not to be underestimated. There is a tight
relationship between the learning algorithm and network structure which makes the
design suitable for the problem [Hay94]. This step ensures the network is compatible
with the problem we are going to solve, as defined by the sample data.
Two different types of neural networks can be distinguished, feed-forward and
recurrent networks. Feed-forward network is a typical neural network consists of
layers, where connections between the units do not form a directed cycle. In a
single layered network there is an input layer and an output layer of neurons. A
multi-layer network has one or more hidden layers of hidden neurons. Extra hidden
neurons increase the ability to extract higher order statistics from data. However,
using too many hidden neurons might leads to overfitting. Figure 3.9 shows the feed
forward network structure. Recurrent network, in the other hand, is a network where
connections between units form a directed cycle and is shown in Figure 3.10. This
structure is believed to be more effective in tasks such as unsegmented connected
handwriting recognition, where they have achieved the best known results [GLF+09].
We select the two-layer feedforward network, with a tan-sigmoid transfer function
in the hidden layer and a linear transfer function in the output layer because this
26Chapter 3. Automatic discovery of connections between Vietnamese’s
anthropometric features
Figure 3.9: A feed-forward network with a single output layer (a) and with onehidden layer and one output layer (b)
Figure 3.10: A recurrent network with hidden neurons
structure can represent any functional relationship between inputs and outputs if
the hidden layer has enough neurons [HDB96]. The design of this neural network
structure is shown in Figure 3.11.
3.4.2 Initialize and train the network
Before training the network, the weights and biases are randomly initialized. The
performance is different each training because of this randomly initiation. The
training process requires a set of inputs p and targets t and begins afterward. This
process tunes the values of the network’s weights and biases to optimize network
performance defined by MSE function. MSE between the network outputs a and
3.4. Discovery of anthropometric relationships using neural networks 27
Figure 3.11: Neural network structure used in the study
target outputs t is defined as follows
mse =1
N
N∑i=1
(ei)2 =
1
N
N∑i=1
(ti − ai)2 (3.2)
In order to train the network, any optimization algorithm can be used to optimize
the performance function. However, some algorithms are believed to have better
performance. These methods use the gradient of the network performance with
respect to the network weights. The gradient is calculated using the backpropagation
algorithm which is an efficient way to calculate the partial derivatives of the network
error function with respect to the weights [Gro02].
There are many training algorithms which make use of the gradients’ information
supplied by the backpropagation algorithm. In these algorithms, a weight update
from iteration k to k + 1 may look like
wk+1 = wk + η.dk (3.3)
where dk is the search direction and η is the learning rate. The training algorithms
are different in ways of determining the search direction and the learning rate.
Different algorithms might also generate different performances.
The fastest training functions are Levenberg-Marquardt function and Quasi-
Newton function. However, these two methods are less efficient for large networks
due to their huge resource consumption. In these cases, Scaled Conjugate Gradient
28Chapter 3. Automatic discovery of connections between Vietnamese’s
anthropometric features
function and Resilient Backpropagation function are better choices. Since our net-
work model is small and Levenberg-Marquardt function performs best on nonlinear
regression problems, we decide to choose this function as our training function.
For each thickness in output, we need to obtain a prediction neural network
model. This can be done by let this thickness be target, and all the input data be
input for the training process. However, as most fields in input do not have any
relationship with the output thickness, the performance might be very bad. We
apply a simple method to increase the performance. We start the training process
with all the input data. We train the network and record the performance over the
validation set. After that, we try removing one field in input data and retrain the
network. If the performance over the validation set this time is worse, we return
the removed field. Otherwise, the removed input field stays outside. We continue
this process until all input field is tried. By this time, we have the set of good
relationship input data with the output thickness, and the model that contains this
relationship.
Chapter 4
Evaluation and Result
We perform the evaluation on the dataset of males which contains 98 samples. In
our evaluation, we use the ten-fold cross-validation to compute the output’s MSE
for the two approaches, linear regression and neural network. As for neural network,
the training is done several times, with the number of neurons from 10 to 20 and
randomly initialized weights each time. The network with best performance over
the validation set is chosen to generate output for the test set. We then compare
these MSE with the ’average method’ in which the output thickness for all tests is
simply the average of all the output in training set. This ’average method’ is what
is used in almost every facial reconstruction systems so far.
Table 4.1 shows our result and their comparisons with the average. It can be seen
from the table that the linear regression always give better result than the average.
Most of the time, neural networks generate the best result over all. However, there
are cases when neural network gives even worse result than average such as result
for zygomatic arch (R), zygomatic (L), gonion (L), and nose height.
In order to deeply analysis, we try plotting results for some random output.
Figure 4.1, 4.2, 4.3, 4.4, 4.5, and 4.6 shows the experiment result. In these figures,
predicted distances are plotted against the true value. For a perfect prediction, the
data should fall along a 45 degree line (the Y=T line), where the outputs are equal
to the targets. The neural network’s values for pronasale thickness, nose length,
pupil-pupil distance are close to the diagonal, indicating the prediction was good.
For linear regression, prediction for nose length and pupil-pupil distance seems to
have good performance. The other predictions are not as good, but acceptable.
29
30 Chapter 4. Evaluation and Result
Table 4.1: MSE values for ’average method’ (AVG), Linear Regression (LR), andNeural Network (NN). The best performance is in boldface.
N# Output AVG LR NN1 vertex 1.1914 1.0625 0.89282 trichion 1.2945 1.0877 1.06643 glabella 1.2074 1.0110 1.07064 nasion 0.9699 0.7571 0.72205 rhinion 0.3886 0.3400 0.37976 pronasale 7.9621 6.0558 5.24567 nose length 21.8621 10.8344 8.70598 subnasale 6.3008 4.3927 4.68789 upper lip border 4.9468 4.3581 3.720510 lower lip border 3.1674 2.7312 2.416711 stomion 2.2193 1.8766 1.816812 metal 4.1007 3.4298 3.362513 meton 2.3685 1.9901 2.088514 opisthooranion 1.8909 1.5124 1.100115 exocanthion (R) 0.7884 0.6635 0.708416 exocanthion (L) 0.8609 0.7121 0.845917 endocanition (R) 2.5804 2.0950 1.721318 endocantion (L) 2.6779 2.0706 2.009919 pupil-pupil 10.8380 4.4587 4.968720 supraobital (R) 0.6689 0.5533 0.455621 supraobital (L) 0.6859 0.5340 0.498622 infraobital (R) 1.4038 1.2479 1.047523 infraobital (L) 1.1147 0.9573 1.192024 zygomatic arch (R) 0.8485 0.7432 1.680525 zygomatic arch (L) 0.8857 0.7400 0.798226 zygomatic (R) 0.8326 0.6982 0.563527 zygomatic (L) 0.9557 0.7722 1.372928 porion (R) 3.3546 2.7241 2.978629 porion (L) 2.5552 2.0471 1.736730 gonion (R) 1.0521 0.9333 0.824531 gonion (L) 0.9360 0.8330 1.544332 alare (R) 2.0965 1.6396 1.593433 alare (L) 2.0342 1.5304 1.449434 lateral nasal (R) 1.9751 1.4220 1.554135 lateral nasal (L) 2.0908 1.3537 1.349536 nose height 4.1012 3.5995 4.568737 bucal (R) 13.6992 11.2034 12.283738 bucal (L) 13.9451 11.6959 11.7598
31
(a) Neural network model(MSE=5.2456)
(b) Linear regressionmodel (MSE=6.0558)
Figure 4.1: Regression results obtained by ten-fold cross validation for pronasalethickness using (a) neural network model and (b) linear regression model.
(a) Neural network model(MSE=8.7059)
(b) Linear regressionmodel (MSE=10.8344)
Figure 4.2: Regression results obtained by ten-fold cross validation for nose lengthusing (a) neural network model and (b) linear regression model.
(a) Neural network model
(MSE=3.7205)
(b) Linear regression
model (MSE=4.3581)
Figure 4.5: Regression results obtained by ten-fold cross validation for upper lipborder using (a) neural network model and (b) linear regression model.
32 Chapter 4. Evaluation and Result
(a) Neural network model(MSE=4.5687)
(b) Linear regressionmodel (MSE=3.5995)
Figure 4.3: Regression results obtained by ten-fold cross validation for nose heightusing (a) neural network model and (b) linear regression model.
(a) Neural network model(MSE=4.9687)
(b) Linear regressionmodel (MSE=4.4587)
Figure 4.4: Regression results obtained by ten-fold cross validation for pupil-pupildistance using (a) neural network model and (b) linear regression model.
(a) Neural network model
(MSE=2.4167)
(b) Linear regression
model (MSE=2.7312)
Figure 4.6: Regression results obtained by ten-fold cross validation for lower lipborder using (a) neural network model and (b) linear regression model.
33
Figure 4.7: Facial Reconstruction Result Using Linear Regression Equations
Figure 4.8: Matching the face and the skull
A complete linear equation for one to one correlation between input and output
is shown in Table 4.2. These equations are used in our facial reconstruction system.
A visual result of our work is given in Figure 4.7. In this figure, the face on the left
is the result of facial reconstruction from the skull in the right. The facial landmarks
are also shown in the skull. Figure 4.8 shows how the face and skull are matched.
34 Chapter 4. Evaluation and Result
Table 4.2: Equations for linear correlation between input and output, with thecorresponding MSE when applied with the whole data set. In the equations, x isthe input and y is the output.
N# Output Input Linear Equation MSE1 vertex cranial breadth y = -0.038041x + 10.8251 1.12302 trichion cranial height y = 0.07284x - 4.6447 1.15003 glabella forehead width (ft-ft) y = 0.073272x - 2.2482 1.04564 nasion cranial height y = 0.070439x - 5.022 0.83005 rhinion molar-molar y = -0.036784x + 4.1817 0.36356 pronasale base nose length y = 0.34191x + 5.6906 6.76847 nose length n-rh y = 1.1274x + 27.5733 11.80508 subnasale bn-bn y = -0.3371x + 22.5646 4.71019 upper lip border molar-molar y = 0.12137x + 5.0732 4.564410 lower lip border cranial height y = 0.086747x + 1.7837 2.947911 stomion base facial length(ba-pr) y = -0.072432x + 10.843 2.007312 metal al-al y = 0.18113x + 3.5632 3.730713 meton al-al y = 0.13342x + 1.0783 2.168414 opisthooranion al-al y = 0.14536x - 0.019573 1.585315 exocanthion (R) cranial height y = 0.052174x - 3.2308 0.707216 exocanthion (L) cranial height y = 0.054596x - 3.4705 0.760917 endocanition (R) cranial height y = 0.10877x - 8.8266 2.285518 endocantion (L) cranial height y = 0.13074x - 11.5992 2.191819 pupil-pupil ex-ex y = 0.71282x - 5.9472 4.631720 supraobital (R) en-en y = 0.083979x + 2.6781 0.588021 supraobital (L) al-al y = 0.097903x + 1.0674 0.586122 infraobital (R) cranial height y = 0.051155x - 2.4011 1.294323 infraobital (L) cranial height y = 0.055695x - 2.9922 1.030924 zygomatic arch (R) n-rh y = 0.070673x + 3.49 0.801325 zygomatic arch (L) base cranial length (n-ba) y = -0.049215x + 9.3495 0.797326 zygomatic (R) base facial length(ba-pr) y = -0.042995x + 8.556 0.759627 zygomatic (L) base facial length(ba-pr) y = -0.051253x + 9.3569 0.838628 porion (R) bn-bn y = 0.1512x + 4.5823 3.031029 porion (L) bn-bn y = 0.16467x + 4.1336 2.170230 gonion (R) al-al y = 0.088104x + 0.53022 0.968431 gonion (L) nasal projection y = 0.10235x + 3.5032 0.884532 alare (R) al-al y = 0.18611x + 1.0906 1.707133 alare (L) al-al y = 0.2063x + 0.33053 1.600034 lateral nasal (R) al-al y = 0.23816x - 2.0783 1.415235 lateral nasal (L) al-al y = 0.2648x - 3.1083 1.384536 nose height base nose length y = 0.17297x + 15.4149 3.750637 bucal (R) bn-bn y = 0.31662x + 6.9254 12.224038 bucal (L) bn-bn y = 0.31206x + 7.0209 12.3088
Chapter 5
Conclusions and Future Work
Facial reconstruction is an interesting research field as it helps in many cases. Re-
searchers have been developing facial reconstruction systems to fasten the manual
process and also to produce better results. In this general problem, one of the most
important issues is to determine the soft tissue thicknesses at landmarks on the skull.
However, most facial reconstruction systems neglect this issue and use the average
thickness for simplicity. Our research has pointed out that this ’average method’
has worse performance than our linear regression method in every case, and worse
than our neural network method in most cases. Our research also shows that there
are relationships between the skull shape and the tissue depths and should people
investigate more and more to discover these relationships. However, our research
has some limitation which can be improved to obtain better results. The following
is our future works.
The first possible development is to improve measurement process. As can be
seen from the experiments, our results show good performance for long distances
such as pronasale thickness or nose length, and bad performance for short distances,
due to the error appeared in the measurement process. This is because the longer
the distance, the less effect it receives from measurement error. In addition, the thin
soft tissues do not depend much on the skull shape, or in other words, they do not
have much relationship with the metrics. In addition, to define the landmarks on
the CT images depends much on the skill and judgment of the people who perform
the measurement, although this technique is the most accurate. This method also
requires a lot of time to measure and collect data. We plan to apply image pro-
cessing to automatic discovery of these metrics. This would save a lot of time in
35
36 Chapter 5. Conclusions and Future Work
measurement and might give better accuracy.
Another thing that needs to be noted is that, in 2009, Pascal Paysan et. al. [PLA+09]
proposed a method to reconstruct the face from the skull, with the capable of tun-
ing the weight and age attributes. From this research, we know that weight and
age affect the facial shape greatly. Our candidates’ age and weight are within wide
range of 18 to 82 and 43kg to 75kg, respectively. Separating candidates into groups
is very important because the relationship between features is different from this
age and weight range to the others and missing this step will lead to moderate error
in training and validation. However, in our experiment, we could not separate the
candidates into groups because the number of entries was not sufficient. Separating
would give even worse result. In the future, we will collect more data for each group
of weight and age. This will improve the prediction performance significantly.
In addition, because our data and problem is straight forward, many other ma-
chine learning techniques can be applied such as the decision stump, support vector
machines, or boosting. With satisfactory results from neural network approach, it
is possibly that better result can be obtained from other techniques. We plan to
implement and analyze result using different techniques.
Lastly, as different landmark configurations might lead to different results and
performances, using different landmark configuration is a worth trying work. This
work requires additional data obtaining from CT images, however.
Publications list
Quang Huy Dinh, Thi Chau Ma, The Duy Bui, Trong Toan Nguyen, Dinh
Tu Nguyen (2011). Facial soft tissue thicknesses prediction using anthropo-
metric distances. Studies in Computational Intelligence, Springer. Proceed-
ings of the 3rd Asian Conference on Intelligent Information and Database
Systems. 2011 (to appear).
37
Bibliography
[Arc97] Katrina Marie Archer. Craniofacial reconstruction using hierarchical
b-spline interpolation. The University of British Columbia, 1997.
[Bul99] David William Bullock. Computer Assisted 3D Craniofacial Recon-
struction. The University of British Columbia, 1999.
[Cai00] Matthew James Francis Cairns. An Investigation into the use of 3D
Computer Graphics for Forensic Facial Reconstruction. Glasgow Uni-
versity, 2000.
[DGPV+06] S. De Greef, Claes. P., D. Vandermeulen, W. Mollemans, P. Suetens,
and G. Willems. Large-scale in-vivo caucasian facial soft tissue thick-
ness database for craniofacial reconstruction. Journal of Forensic Sci-
ences, 159:126–146, 2006.
[Dum86] E. R. Dumont. Mid-facial tissue depths of white children: An aid to
facial feature reconstruction. J Forensic Sci, 1986.
[EMS01] I.H. El-Mehallawi and E. M. Soliman. Ultrasonic assessment of facial
soft tissue thicknesses in adult egyptians. Journal of Forensic Sciences,
117(1-2):99–107, 2001.
[Geo87] R. M. George. The lateral craniographic method of facial reconstruc-
tion. Journal of Forensic Sciences, 32(5):1305–1330, 1987.
[GLF+09] A. Graves, M. Liwicki, S. Fernandez, R. Bertolami, H. Bunke, and
J. Schmidhuber. A novel connectionist system for improved uncon-
strained handwriting recognition. IEEE Transactions on Pattern Anal-
ysis and Machine Intelligence, 31(5), 2009.
38
Bibliography 39
[Gro02] R. Grothmann. Multi-Agent Market Modeling based on Neural Net-
works. Ph.D. thesis, Faculty of Economics, University of Bremen, Ger-
many, 2002.
[Hay94] S. Haykin. Neural Networks. A Comprehensive Foundation. Macmillan
College Publishing, New York, 1994.
[HDB96] Martin T. Hagan, Howard B. Demuth, and Mark H. Beale. Neural
Network Design. PWS Publishing Company, Boston, Massachusetts,
1996.
[HLW85] L. S. Hodson, Lieberman, and P. Wright. In vivo measurement of facial
thickness in american caucasoid children. J Forensic Sci, 1985.
[Jon01] Mark W. Jones. Facial reconstruction using volumetric data. Proceed-
ings of the Vision Modeling and Visualization Conference 2001, pages
135–150, November 2001.
[KHS03] Kolja Kahler, Jorg Haber, and Hans-Peter Seidel. Reanimating the
dead: reconstruction of expressive faces from skull data. ACM Trans-
actions on Graphics (TOG), 22(3):554561, 2003.
[Kro46] Wilton Marion Krogman. The reconstruction of the living head from
the skull. FBI Law Enforcement Bulletin, 15(7):1–8, July 1946.
[MLB+00] M. H. Manhein, G. A. Listi, R. E. Barsley, Musselman R., N. E. Bar-
row, and D. H. Ubelaker. In vivo facial tissue depth measurements for
children and adults. Journal of Forensic Sciences, 45(1):48–60, 2000.
[PLA+09] Pascal Paysan, Marcel Luthi, Thomas Albrecht, Anita Lerch, Brian
Amberg, Francesco Santini, and Thomas Vetter. Face Reconstruction
from Skull Shapes and Physical Attributes. 5748:232–241, 2009.
[PS96] V. M. Phillips and N. A. Smuts. Facial reconstruction: Utilization of
computerized tomography to measure facial tissue thickness in a mixed
racial population. Forensic Sci Int., 83:51–59, 1996.
[Rhi84] Stanley Rhine. Tissue thickness measures: American caucasoids, amer-
ican blacks, southwestern indians. Physical Anthropology Laboratories,
Maxwell Museum of Anthropology, University of New Mexico, 1984.
40 Bibliography
[SC10] C. N. Stephan and J. Cicolini. Tallied facial soft tissue depth data
(tfstdd), 2010.
[SJG+02] D. Sahni, I. Jit, M. Gupta, P. Singh, S. Suri, Sanjeev, and H. Kaur.
Preliminary study on facial soft tissue thickness by magnetic resonance
imaging in northwest indians. Forensic Science Communications, 4,
2002.
[VPST+07] J. Vander Pluym, W. W. Shan, Z. Taher, C. Beaulieu, C. Plewes, A. E.
Peterson, O. B. Beattie, and J. S. Bamforth. Use of magnetic resonance
imaging to measure facial soft tissue depth. Cleft Palate-Craniofacial
Journal, 44:52–57, 2007.
[Wel83] H. Welcker. Schiller’s schdel und todenmaske, nebst mittheilungen ber
schdel und todenmaske kants. 1883.
[Wil04] Caroline Wilkinson. Forensic facial reconstruction. Cambridge Univer-
sity Press, 2004.