FUZZY SKIN DETECTION

FUZZY SKIN DETECTION









TEY YI CHIN









A project report submitted in partial fulfillment of the

requirements for the award of the degree of

Master of Science (Computer Science)









Faculty of Computer Science and Information System



Universiti Teknologi Malaysia









OCTOBER 2008

iii









To my beloved father and mother,



Tey Kim Seng and Chia Ah Mui

iv









ACKNOWLEDGEMENT









During the preparation of this project, many people have been very kind to

lend a hand to help and giving advices. Their advices and help have been very

valuable and largely contributed for me to conduct this project. Taking the chances

here, I would like to express my appreciation to all of them.







First of all, I wish to send my gratitude to my supervisor, Assoc. Prof. Dr. Ali

Selamat. His invaluable experience and advice has been a light in the dark to me

throughout the project. It was my pleasure to be under his guidance. I am also very

thankful to my co-supervisor, Prof. Dr Mohd Aizaini Maarof for his encouragement

and guidance during the project conduction. His guidance has been a valuable asset

for me. At the same time, I would like to thank Assoc. Prof. Dr Siti Mariyam for

sharing her knowledge with me.







Nonetheless, I would like to send my gratitude to FSKSM postgraduate staffs,

especially Pn. Lijah. They have been very helpful throughout my master study.

Besides that, the supports and helps from my fellow course mates and friends should

not be forgotten. They have been very helpful and sharing their ideas and knowledge

with me. Their encouragement and assistance is my motivation throughout the

project implementation. Due to the limit space, I apologize for not being able to list

them all down here.







Finally, I would like to send my appreciation to my parents for their support

and care throughout my whole life.

v









ABSTRACT









The robustness of web technologies allows huge collection of information to

be available through internet. Among this information, some of them are unhealthy

and undesirable. This kind of harmful information will make bad impact on human

development especially to children and teenagers. Therefore, an explicit image

filtering method based on skin detection is proposed in this project. Skin detection is

a popular image processing technique that has been applied in many areas such as

video-surveillance, cyber-crime prosecution and face detection. It is well known

technique to detect the human appearance within image. However, it faces several

drawbacks when using color as cue to detect skin. First, the similarity between skin

and background color within an image. Second, the skin appearance of human under

different lightning condition also adds the complexity in skin detection. Third,

different camera characteristics also influence the performance of skin detection.

Therefore in this project, fuzzy theory is proposed to improve the skin detection

performance by solving the first problem. By improving the first problem, we can

increase the classification accuracy when discriminate human and animal skin

images which was tested in this project. The complexity of applying fuzzy theory in

skin detection is based on the highly similarity between skin and non-skin pixels as a

clear description of the skin color set is needed for processes. Hence, the rules need

to be explicitly enough in order to achieve better performances. Finally, experiment

has been conducted to test the applicability of fuzzy classification. Although the

classification result is lower than comparison method, fuzzy theory has been proved

to be able to discriminate human and animal skin.

vi









ABSTRAK









Perkembangan teknologi web membenarkan maklumat yang berjumlah besar

dihantar melalui. Antara maklumat yang dihantar antara web ini, sebahagian adalah

dalam bentuk gambar yang tidak sihat. Gambar-gambar ini menjejaskan kesihatan

dan perkembangan manusia, terutamanya kepada kanak-kanak dan remaja. Oleh itu,

satu kaedah untuk menapiskan gambar-gambar tidak sihat diperkenalkan dalam

projek ini. Teknik pengesahan kulit manusia melalui gambar merupakan salah satu

teknik popular. Teknik ini digunakan dalam banyak bidang seperti video pengawasan

dan pendakwaan jenayah siber. Bagaimanapun, beberapa masalah wujud semasa

menggunakan teknik ini. Masalah pertama yang ditemui ialah kesamaan antara

warna kulit manusia dengan warna latar belakang gambar. Selain itu, warna kulit

manusia berbeza apabila dalam keadaan cahaya yang berbeza. Pada masa yang sama,

ciri-ciri kamera yang berlainan juga akan mempengaruhi prestasi ketepatan untuk

mengesahkan kewujudan kulit manusia dalam gambar. Dalam projek ini, Teori

Fuzzy telah digunakan untuk menyelesaikan masalah pertama. Selain itu, Teori

Fuzzy juga digunakan untuk membezakan gambar manusia dengan gambar haiwan

dalam projeck ini. Kesusahan untuk melaksanakan Teori Fuzzy dalam projek ini

ialah membezakan warna kulit manusia dengan warna kulit haiwan yang mempunya

nilai serupa. Peraturan Fuzzy yang jelas diperlukan untuk mencapai prestasi yang

lebih baik semasa mengesahkan kewujudan kulit manusia dalam gambar. Walaupun

prestasi ujikaji adalah lebih buruk daripada kaedah dibandingkan, Teori Fuzzy telah

dibuktikan dapat membezakan gambar kulit manusia dengan gambar haiwan.

vii









TABLE OF CONTENTS









CHAPTER TITLE PAGE





TITLE PAGE i

DECLARATION ii



DEDICATION iii

iv

ACKNOWLEDGEMENTS

ABSTRACT v

ABSTRAK vi

TABLE OF CONTENTS vii



LIST OF TABLES xi

LIST OF FIGURES xii



LIST OF ABBREVIATIONS xiv



LIST OF SYMBOLS xv

LIST OF APPENDIX xvi





1 INTRODUCTION 1

1.1 Introduction 1

1.1.1 Application of Skin Detection In Video 2

Surveillance

1.1.2 Application of Skin Detection In Hand 2

Detection

1.1.3 Application of Skin Detection In Cyber- 2

crime Prosecution

1.2 Problem Background 3

1.3 Problem Statement 5

1.4 Project Aim 5

1.5 Objectives 6

viii





1.6 Project Scopes 6

1.7 Significant of Study 7

1.8 Conclusion 7





2 LITERATURE REVIEW 8

2.1 Introduction 8

2.2 Colorspaces 9

2.2.1 RGB 9

2.2.2 Normalized RGB 10

2.2.3 HIS, HSV, HSL – Hue Saturation 11

Intensity (Value, Lightness)

12

2.2.4 TSL – Tint, Saturation, Lightness

2.2.5 YCbCr 13



2.3 Skin Modeling 16

2.3.1 Explicitly Defined Skin Region 16

2.3.2 Non-parametric Skin Distribution 17

Modeling

2.3.2.1 Normalized Lookup Table 17

(LUT)

2.3.2.2 Bayes Classifier 18

2.3.3 Parametric Skin Modeling 20

2.3.3.1 Single Gaussians 21

2.3.3.2 Mixture of Gaussian (MoG) 21

2.3.3.3 Elliptical Boundary Model 22

2.4 Fuzzy Methods 25

2.4.1 Takagi Sugeno Fuzzy Inference 25

System

2.4.2 Modified Fuzzy C-Mean Algorithm 26

(MFCM)

2.4.3 Fuzzy Fusion 27

2.4.4 Linear Matrix Inequality (LMI) 29

Fuzzy Clustering



2.5 Summary 33

ix





3 METHODOLOGY 34

3.1 Introduction 34

3.2 Proposed Methodology 35

3.2.1 Data collection 38

3.2.2 Preprocessing 38

3.2.3 Feature Extraction 39

3.2.3.1 Colorspace Selection 39

3.2.3.2 Skin Modelling 39

3.2.3.2.1 Integration of Skin 41

Modelling in Fuzzy

Rules

3.2.4 Background Subtraction 42

3.2.5 Fuzzy Classification 44

3.3 Result Evaluation Method 47

3.4 Summary 49





4 RESULT AND DISCUSSION 50

4.1 Introduction 50

4.2 Skin Detection in Images 50

4.3 Experimental Setup 52

4.3.1 Skin Modelling 52

4.3.2 Background Subtraction 54

4.3.3 Fuzzy Classification 54

4.4 Experimental Result 55

4.4.1 Analysis of Fuzzy Logic Ability to 56

Discriminate Human and Animal

Skin Images

4.4.2 Analysis of Fuzzy Logic Ability to 57

Discriminate Porn and Non-Porn

Images

4.5 Result Comparison 57

x





4.5.1 Results Comparison on Human and 59

Animal Images Classification

between Fuzzy Logic, Explicitly

Defined Skin Region and Takagi

Sugeno Fuzzy Inference System

4.5.2 Results Comparison on Porn and 60

Non-Porn Images Classification

between Proposed Method,

Explicitly Defined Skin Region and

Takagi Sugeno Fuzzy Inference

System

4.5.3 Accuracy Rates Comparison between 61

Proposed Method, Explicitly Defined

Skin Region and Fuzzy Sugeno

Inference System on Different

Images Type

4.6 Result Discussion 62

4.7 Conclusion 65





5 CONCLUSION 66

5.1 Introduction 66

5.2 Project Achievement 67

5.3 Future Work 68

5.4 Conclusion 68





REFERENCES 69

APPENDIX A 72

APPENDIX B 73

APPENDIX C 74

xi









LIST OF TABLES









TABLE NO TITLE PAGE





2.1 Comparison of colorspace 14

2.2 Comparison of classifier 23

2.3 Comparison of fuzzy method on skin detection 32

3.1 Modules in methodology 37

3.2 Explanation of data collection 38

3.3 Explanation of preprocessing 39

3.4 Explanation of feature extraction 40

3.5 Explanation of Background Subtraction 44

3.6 Explanation of Fuzzy Classification 47

4.1 Skin Color Boundary of RGB colorspace 53

C.1 Classification Results of Human and Animal Images 74

C.2 Classification Result of Porn and Non-Porn Images 75

xii









LIST OF FIGURES









FIGURE NO TITLE PAGE





2.1 RGB colorspace color cube 10

2.2 HSV graphical depiction 11

2.3 TSL colorspace 12

2.4 YCbCr colorspace 13

2.5 Application of explicit defined skin region in finding 17

face

2.6 3D color histogram 18

2.7 The skin/non-skin u-matrix. 20

2.8(a) Takagi Sugeno fuzzy inference system input image 26

2.8(b) Takagi Sugeno fuzzy inference system output image 26

2.9 Block diagram of the framework for the detection of 29

skin based on the fusion paradigm

3.1 General skin detection process 35

3.2 Methodology of this project 36

3.3 Integration of skin modeling in membership function 41

4.1 Experiment Method 51

4.2 Skin discrete probability of red color channel 52

4.3 Skin discrete probability of green color channel 53

4.4 Skin discrete probability of blue color channel 53

4.5(a) Image before background subtraction 54

4.5(b) Image after background subtraction 54

4.6 Mamdani Fuzzy Inference Method 55

4.7 Receivers Operating Characteristic Curve (ROC) of 56

skin percentage of human and animal skin images

samples

xiii





4.8 Receivers Operating Characteristic Curve (ROC) of 58

skin percentage of porn and non-porn images samples

4.9 ROC Graph on comparison of classification result on 59

human and animal images

4.10 ROC Graph on comparison of classification result on 60

porn and non-porn images

4.11 Comparison of accuracy rates of different methods on 61

different image types

4.12(a) Image without losing data before subtraction 62

4.12(b) Image losing data after subtraction. 62

4.13 Comparison of accuracy before and after subtraction 63

4.14 Image with bright background 63

4.15(a) Human skin images falsely detected as false positive 64

4.15(b) Human skin images falsely detected as false positive 64

xiv









LIST OF ABBREVATIONS









B - Blue

BG - Blue Green

Cb - Chrominance Blue

CIE - Commission On Illumination

Cr - Chrominance Red

CRT - Catorade Ray Tube

EM - Expectation Maximazation

FPR - False Positive Rate

G - Green

HIS - Hue Saturation Intensity

HSL - Hue Saturation Lightness

HSV - Hue Saturation Value

LUT - Normalized Lookup Table

ML - Bayes Maximum Likelihood

MoG - Mixture of Gaussian

pdf - Probability Joint Function

R - Red

RB - Red Blue

RG - Red Green

ROC - Receiver Operating Characteristic

SGM - Symmetric Gaussian Model

SOM - Self Organizing Map

TPR - True Positive Rate

TSL - Tint, Saturation, Lightness

Y - Luma

xv









LIST OF SYMBOLS









c - Color vector

k - Prior probabilities

N - Number of mixture components

Norm - Normalized coefiecient either sum of all histogram

value or maximum bin value

n(c) - Pixel counts in the c-bin color of non-skin class



P (c ) - Probability Distribution



s (c ) - Pixel count in the c- bin color skin class



fi - Number of samples



Tn - Total counts in non-skin color histogram bin



Ts - Total counts in skin color histogram bin



wi - Weighted factor, contribution of ith Gaussians



- Determinant Constant

 - Diagonal covariance matrix

 - Mahalanobis distance

 - Mean Vector

 - Model parameter

 - Model parameter

 - Threshold Value

 - Threshold value chosen as trade off between true and

false positives

xvi









LIST OF APPENDIX









APPENDIX NO TITLE PAGE

1 20 Samples of Human Skin Images 72

2 20 Samples of Animal Skin Images 73

3 Classification Results 74

CHAPTER 1









INTRODUCTION









1.1 Introduction





With fast-growing World Wide Web and more simplified way to access

Internet, people are easier and have more choices to find and locate information.

Many of this information available in Internet are in the form of images such as

photo, videos etc. As most people want to understand the image contents before they

locate it, the importance of image processing increases. So, various image processing

methods have been proposed to achieve this objective. One of the important

techniques of image processing is the skin detection technique. As a popular

technique, it proposes a way to detect human and human-body parts in images using

skin value. It has been used in various applications such as video surveillance,

human-machine interface, cyber-crime prosecution, gesture recognition, hand

detection, teleconference. The application of skin detection in video surveillance,

hand detection and cyber-crime prosecution will be discussed in Section 1.1.1,

Section 1.1.2 and Section 1.1.3.

2





1.1.1 Application of Skin Detection in Video Surveillance





The application of video surveillance allows users to identify, alert, analyse

and respond to the threat such as theft and terrorism. This is where the skin detection

comes in as it can be used to detect the face of the people in the video who may be

the person involve in stealing if theft happened or the culprit that commit a criminal.

It provides help for police to deal with criminals and find the culprits. Various

approaches have been proposed by researchers to improve the application of skin

detection in video surveillance. T. Kim et al (2005) had used an integrated approach

of multiple face detection techniques to improve the performance of skin detection in

video surveillance. Then, Robertson et al (2006) used the skin detection method to

estimate the direction of the head in order to help to determine human behaviour.









1.1.2 Application of Skin Detection in Hand Detection





The hand detection is implemented to measure the length and shape of fingers

and knuckles. Optical camera and light-emitting diodes that have mirrors and

reflectors are used to take two orthogonal, two dimensional images of the back and

side of hands. Skin detection technique is applied in hand detection to improve the

performance of the hand detection. Based on the research by K¨olsch and Turk

(2004), the best detection they obtained achieved outstanding performance in

practical application, indoors and outdoors with the combination of skin color

verification.









1.1.3 Application of Skin Detection in Cyber-crime Prosecution





Although the robustness of information available within the Internet provides

people a better way to acquire information, it also exposes people to the illegal and

harmful content such as violent and nude image. There is a need to block and filter

this undesirable content from the web especially from children. To fulfil this

3





requirement, some companies such as NetNanny and SuftWatch operate by

maintaining lists of URL’s and newsgroups and require constant manual updating.

Due to the rapid changing and growing of Internet, this approach is not enough to

cope with all these evolving web contents. By taking advantage of the strong

relationships between adult images and images with large patch of skin, a skin

detector can be developed in order to provide an efficient and effective way to detect

adult images. Several researches have been done to fit the skin detection in web

content filtering. Zhang et al (2004) suggested that the adult images can be blocked

by using statistical skin detection. They build a first order model which imposes

constraint on color gradients of neighbouring pixels.





The application of skin detection in various fields shows its popularity within

image processing technique. It is computationally effective and robust information

against rotation, scaling and partial occlusions. However, this skin color modeling is

not perfect yet, as there are still problems to be solved to achieve high accuracy and

effective skin detection method.









1.2 Problem Background





Skin detection has been a popular technique to classify or detect the skin

pixels within an image. Several advantages have been provided by skin detection.

First, it is fast in processing due to its low level processing. Then, Stern and Efros

(2002) stated that it is also invariance against rotations, partial occlusions and poses

change. However, a survey conducted by Kakumanu et al (2007) has shown several

factors need to be considered to apply skin detections which are illumination, camera

characteristic and ethnicity.







The illumination factor related with the lightning condition when an image is

taken which including indoor, outdoor, highlights, shadows and non-white lights.

The transformation of lightning condition in an environment will produce a change

4





on the skin color that leads to color constancy problem. This color constancy

problem will seriously affect the performance of the skin detection system







The camera characteristic is difference from one camera to other camera

depending on the camera sensor characteristic. Although under same lightning

condition, the performance of the skin detection system will still be variant depend

on the camera characteristic.







Although the skin color differences which vary from one region to another

region can be identified using human eyes, the differentiation and variation of skin

color becomes difficult when it comes to computation. For example, the skin color of

people between different ethnicity from Caucasian to Asian and African ranges from

white to yellow and black. Besides that, the color value of human skin pixel and

other pixels within an image may fall within the same color boundary which makes

the computation complicated. For example, the color value of lion pixel value may

have similar RGB value as human skin value.







The problems that have been encountered in visual spectrum can be solved by

applying non-visual spectrum such as infrared and spectral imaging which is stated

by Terrillon et al (1998). Skin color in non-visual spectrum method is invariant to

changes in illumination conditions, ethnicity, and make up and shadow. However,

the performance of non-visual spectrum method was decreased by the tiresome setup

procedure and high cost equipments. Therefore, there is still much room to be

improved in the area of skin detection.

5





1.3 Problem Statement







To solve the color likelihood problem between human skin pixels and background

pixels, the skin color boundary need to be determined which can be achieved by

building a skin color model. This identified skin color boundary will be used as

references in fuzzy classification later to classify a pixel as skin or non-skin pixels.

So, this project examines the following questions







How the color likelihood problem of human skin pixel and background pixel can be

solved by building a skin color model? How the fuzzy logic can be used to classify

skin and non-skin pixels based on the identified skin color model?







Then, this project intends to classify human and animal images using Fuzzy Logic.

Therefore, another question is







How fuzzy rules can be used to classify human and animal images? What is the

performance of fuzzy logic when classifying human and animal images?









1.4 Project Aims





This project will use RGB colorspace to extract color features. These

extracted color features will be used to build skin color model. Types of existing

colorspaces include Red Green Blue (RGB), normalized RGB, Hue Saturation

Intensity (HIS) and YCbCr.







In this project, suitable skin color modelling method that works well with

RGB colorspace will be identified. The skin color modelling method will help to

discriminate the human skin and animal skin images. The available skin modelling

methods are explicitly defined skin region, nonparametric skin distribution modeling,

parametric skin distribution modeling and dynamic skin distribution method.

6





The images will be subtracted in order to acquire object of interest. Then, the

subtracted images will be processed with fuzzy method in order to solve problem in

skin detection method. In the end, an algorithm that can classify human skin and

animal images with high true positive rate and low false positive rate is desirable.









1.5 Objectives







The objectives of this project are as follows:

1. To propose a new set of fuzzy rules in discriminate human skin and skin-

like pixels.

2. To study human and animal skin images using fuzzy classification.

3. To evaluate and compare the classification result with simulation result of

explicitly defined skin region and Takagi-Sugeno Fuzzy Inference System









1.6 Project Scopes







This project is focus on the skin color extraction and classification. The

performance of fuzzy classifier is evaluated. This project covers the following area:

1. Asian skin region: Only skin color information of Asian people will be

covered in this project.

2. Image format: Images in JPEG format will be used to perform skin

detection

3. RGB colorspace will be used for feature extraction: This project will use

RGB colorspace to represent the skin color information.

4. Fuzzy classification for skin detection: Fuzzy logic will be used in this

project for the classification of skin and non-skin pixels.

5. This project will be conducted by using vb.net as programming language.

7





1.7 Significance of Study







This study evaluates the performance of skin detection based on fuzzy

classification. The result is compared to the result acquired from previous skin

detection technique to check whether this approach can achieve better classification

performance. The result of the evaluation will help to solve the problems faced when

discriminating human skin pixels and pixels that have similar color information as

human skin. This approach could be used to the development of a methodology that

will be of value in filtering adult images within web.









1.8 Conclusion







In conclusion, the report consists of 5 chapters. Chapter 1 presents the

introduction of the study, problems background, objectives and scopes of the project.

Chapter 2 gives literature reviews on the existing colorspace and skin detection

technique. Project methodology is discussed in Chapter 3 and Chapter 4 exhibits the

experimental results. The conclusion and suggestion for future work will be

explained in Chapter 5.

CHAPTER 2









LITERATURE REVIEWS









2.1 Introduction





Considers that a large part of information that flows through web is images,

the importance of image processing as a tool to understand image contents is highly

increased. Skin detection as one of the popular image processing techniques is used

to track people, filtering web contents and facilitating human-computer interaction.

Skin detection provides several attractive attributes, (i) high processing speed due to

its low level processing and (ii) invariance against rotations, partial occlusions and

pose change.







From the past decades, many methods have been proposed to detect the skin

pixels within the images, photos and videos. Although different methods have been

proposed to conduct skin detection, Kakumanu et al (2007) had stated that it

basically involves two main processes. These two important processes are feature

extraction and classification. In skin detection, the feature extraction process

involves extracting color features from images. The rationale behind this approach is

that human skin has very consistent color which distinct from other object. When

using color as cue to detect skin, a question appeared: which colorspace better

represent skin color. Many researchers have centered their work on selecting the

suitable and best colorspace to work on skin classification method which is first step

in building a skin color modeling system. Therefore, the existing colorspace used to

represent color information will be discussed in Section 2.2. Among the existing

colorspaces are RGB, normalized RGB, HIS, HSV, YCbCr and YIQ. In this chapter,

9





we discuss the advantages and disadvantages of some of the most popular

colorspaces employed.







After feature extraction, these extracted features will be classified to

determine whether the particular pixel or region belongs to skin color. To achieve

this intention, many modelling techniques have been proposed. Among the proposed

techniques are explicitly defined skin region, non-parametric skin distribution

modeling, parametric skin-color modeling. These stated techniques will be further

discussed in Section 2.3. Besides that, fuzzy methods that have been used in skin

detection are discussed in Section 2.4. Finally, the summary of this chapter will be

given in Section 2.5.









2.2 Colorspaces









2.2.1 RGB







As one of the most common used colorspace, it is widely used for processing

and storing of digital images. RGB corresponds to the combination of three primary

colors, Red, Green and Blue. It is originally developed with CRT as an additive

colorspace and used to extract color texture features separately or from cross-

correlation between RB, RG, BG planes. Jones and Rehg (1999) have used the RGB

colorspace to construct a Histogram Color Model to identify the separability of skin

and non-skin pixels. However, several disadvantages have limited the usefulness of

this colorspace, especially in color analysis and color-based algorithms. These

disadvantages include the high correlation between channels, significant perceptual

non-uniformity, mixing or chrominance and luminance data. The color cube of RGB

colorspace is shown in Figure 2.1.

10









Figure 2.1: RGB colorspace color cube









2.2.2 Normalized RGB







To reduce the effect of lighting condition and ethnicity towards the

performance of skin detection, RGB colorspace is normalized by using a simple

procedure which is shown in (1).







R G B

r , g , b

RG B RG B RG B (1)







As a result of normalization, the sum of the normalized components is unity (r + g +

b = 1) and less dependence on lighting. Since the sum of these components is 1, the

third component does not hold any significant information and can be diminished to

obtain dimensionality reduction. One of the remarkable properties of this colorspace

is that it is invariant to the changes surface orientation for matte and dull surface

relative to lightning condition. It has been used by Mustafa and Abdelazeem (2005)

and Diplaros et al (2004).

11





2.2.3 HIS, HSV, HSL-Hue Saturation Intensity (Value, Lightness)







Hue-saturation based colorspace is another popular colorspace that is based

on human perception color. H stands for Hue which defines the dominant color of an

area. S stands for Saturation which measures the colorfulness of an area in proportion

to its brightness. The intensity, value and lightness are all related to the color

luminance. It was introduced for users who need to define the color properties

numerically. The graphical depiction of HSV colorspace is shown in Figure 2.2.









Figure 2.2: HSV graphical depiction







One of the advantages of Hue-saturation based colorspace is it is invariant to

highlight in white light sources. Besides that, it is also invariant to matte surfaces,

ambient light and surface orientation relative to light sources. However, the bad

confliction between computation of “brightness” and the color vision properties

make it becomes undesirable. It can be transformed to and from RGB colorspace

based on the formulas shown in (2), (3) and (4).







1

(( R  G )  ( R  B))

H  arccos 2 (2)

(( R  G ) 2  ( R  B)(G  B))



min( R, G , B) (3)

S  1 3

RG B

1 (4)

V  ( R  G  B)

3

12





2.2.4 TSL – Tint, Saturation, Lightness







TSL colorspace is similar and closes to Hue and Saturation in meaning, as it

is transformed from of normalized RGB into more intuitive in values. It has been

proved to perform well in face detection and Gaussian Modelling. The

transformation from RGB to TSL is showed in (5), (6) and (7).







 1 r' 1

 2 arctan g '  4 , if g'  0



 1 r' 3 (5)

T   arctan  , if g'  0

 2 g' 4

1

 , if g'  0

2



9 2

S (r '  g ' 2 )

5 (6)

L  0.299 R  0.587G  0.114 B (7)







1 1

where r '  r  , g '  g  and r, g come from (1). Due to the chromaticity, non-

3 3

uniformity is found existing in TSL colorspace even in the condition which camera

was calibrated to the light sources. The TSL colorspace has been used by Terillon et

al. (1998) and Brown et al (2001). Figure 2.3 shows the 3D model of TSL colorspace.









Figure 2.3: TSL colorspace

13





2.2.5 YCrCb







YCrCb is an encoded non-linear RGB signal that used by European television

studios and for image compression. As Y is the luma component, which is the

weighted sum of RGB value while Cr is the red and Cb is the blue chroma

component. The transformation from RGB to YCbCr is shown in (8), (9) and (10).







Y  0.299 R  0.587G  0.114 B (8)

Cr  R  Y (9)



Cb  B  Y (10)







In YCrCb colorspace, intensity component can be easily access. This

colorspace reduces the redundancy present in RGB channels and represent the color

with statistically independent component. Kumar et al (2006) removed the Y and Cb

components from the converted image and only the Cr component remained as it

proves that all skin candidates is rich in red color. It is also adopted in Hsu et al

(2002) because of its perceptually uniform and it is similar to the TSL colorspace in

terms of the separation of luminance and chrominance as well as the compactness of

the skin cluster. Figure 2.4 shows the 3D model of YCbCr colorspace.









Figure 2.4: YCbCr colorspace

14





Table 2.1: Comparison of Colorspace

Colorspace Paper Technique Advantages Disadvantages Justification

Category



2.2.1 Basic A Probabilistic Approach To Skin Detection Explicitly Skin 1. Nullifying brightness 1. High false positive and 1. Non-skin region and skin

Colorspace (F.H.B. Cardoso, 1999 ) Color Boundary discrepancies among negative rates color region may have

(RGB) pixels similar RGB value.





An Approach To Image Extraction And Accurate Explicitly Skin 1. Widely used for 1. Not reliable as represent 1. Skin luminance may vary

Skin Detection From Web Pages (Girgis et al, Color Boundary storing and not only color but within and across person

2007) processing digital luminance due to ambient lightning.

images



Statistical Color Model With Application To Skin Statistical Model 1. High detection rates 1. Fail to detect correctly 1. The false detections are

Detection (Jones and Rehg, 1999) with Histogram in highly saturated or sparse and scattered in

Gaussian Model shadowed skin. non-skin images,

especially images with

wood and copper color

model

2.2.2 Basic Face Detection Based On Skin Color Using Neural Neural Network 1. Contains no 1. Detected skin region is 1. Lose data since luminance

Colorspace Network (Mustafa and Abdelazeem, 2005) luminance discontinuous due to information is eliminated

(Normalized information lightning effect

2. Contain only two

RGB)

components which

speeds up calculation

Skin Detection Using The EM Algorithm With Expectation 1. Skin component is 1. Low performance in 1.3.3 Performance may

Spatial Constraints (Diplaros et al, 2004) Maximization adequately describe detecting saturated or decrease when detection

Algorithm shadow skin shadow region



A SOM Based Approach To Skin Detection With Self Organized 1. Show consistent 1. 5% of Performance 1. The effect of camera

Application In Real Time System (Brown et al, Map result with Gaussian Loss over 20 properties toward

2001) method configurations performance are not

considered.









14

15





Colorspace Paper Technique Advantages Disadvantages Justification

Category

2.2.3 & 2.2.4 Estimation And Prediction OF Evolving Color Normalized 1. Robust to minor 1. Costly to use to convert 1. Inconvenient for

Perceptual Distribution For Skin Segmentation Under Varying Lookup illuminant changes from RGB parametric skin color

Colorspace ( HIS, Illumination (Sigal et al, 2000) Table(LUT) 2. Perform well for models that need tight

HSV,HSL,TSL) estimation and cluster for skin region.

prediction of skin

color distribution

evolution in images

sequences taken from

video

Active Face and Feature Tracking (Jordao et al. Explicitly Defined 1. Skin area is easily 1. Unable to detect people 1. Low performance when

1999) Skin Region detected when it falls with different skin working under

within the threshold region unconstraint environment





Comparative Performance of Different Gaussian Model 1. High detection 1. False positive occurred 1. False positive may be

Chrominance Spaces For Color Segmentation and performance caused by strong shadow

Detection of Human Faces in Complex Scene or strong illumination

Image (Terrillon

et al. 2000)

2.2.5 Orthogonal An Efficient Skin Illumination Compensation Explicitly Defined 1. Intensity component 1. Detection rates decrease 1. Losing image data as

Colorspace Model For Efficient Face Detection (Ravi et al, Skin Region can be easily when face merged with lightning information is

2006) accessed background color compensated





Face Detection On Color Image (Hsu et al. 2002) Gaussian Model 1. Perceptually uniform 1. Face difficulty in 1. Losing image data as

and widely used in detecting low-luma and lightning information is

video compression high-luma skin tone compensated

method.

2. Similar to TSL in

terms of separation of

luminance and

chrominance









15

16





2.3 Skin Modeling







Kakumanu et al (2007) stated that there are basically two types of skin

modelling methods which are pixel and non-pixel based classification. A lot of

different techniques and approaches have been used by different researches to duel

with the problems. These techniques are usually used to measure the distance of the

pixel color to the skin tone.









2.3.1 Explicitly Defined Skin Region







This is one of the easiest methods as it explicitly defines skin-color

boundaries in different colorspaces. Different ranges of thresholds are defined

according to each colorspace components as the image pixels that fall between the

predefined ranges are considered as skin pixels. A performance comparison of three

different colorspaces, CbCr, HS and GT based on explicitly defined skin region has

been stated in Phung et al (2005). The advantage of this method is obviously the

simplicity which normally avoids of attempting too complex rules to prevent over-

fitting data. However, it is important to select good colorspace and suitable decision

rules to achieve high recognition rate with this method. The application of this

method can be seen in Gomez and Morales (2002), Phung et al (2005), Girgis et al

(2005) and Peer et al (2003). The example of explicit defined skin region which is

implemented in Peer et al (2003) was shown in (11). The application of this method

in face recognition is shown in Figure 2.5.







(R, G, B) is classified as skin if

R > 95 and G >40 and B >20 and

Max{R, G, B} – min{R, G, B} > 15 and (11)

|R - G| >15 and R >G and R > B

17









Figure 2.5: Application of explicit defined skin region in finding face. 1) Reduces the

original image resolution of 2048×1536 pixels to 160×120 pixels. 2) Eliminate all

pixels that cannot represent a face. 3) Segment all the regions containing face-like

pixels using region growing. 4) Eliminate regions, which cannot represent a face

using heuristic rules.









2.3.2 Non-parametric Skin Distribution Modeling







Non-parametric skin distribution modeling estimates skin color distribution

from the training data without deriving an explicit model of skin color.









2.3.2.1 Normalized Lookup Table (LUT)







In Lookup table, the chrominance plane of colorspaces is quantized into

number of bins which corresponding to particular range of color component value

pairs (in 2D case) and triple (in 3D case) to form a 2D or 3D histogram. It is a stable

18





object representation unaffected by occlusion changes in the view and can be used to

differentiate a large number of objects. The histogram count is then normalized,

converting histogram value into discrete probability distribution:







skin[c]

P skin  (12)

Norm







where skin [c] gives the value of the histogram related to the color c and the Norm is

the normalized coefiecient either sum of all histogram value or maximum bin value.

The lookup table then gives the normalized value which estimates the possibility that

the color in that pixels belong to skin. Figure 2.4 shows the 3D color histogram built

from RGB colorspace which is used by Jones and Rehg [5].









Figure 2.6: 3D color histogram









2.3.2.2 Bayes Classifier







The value of Pskin (c ) obtained in the normalization is further developed to two



different types of histogram in Jones and Rehg [5]. It is defined as the probability

that an observed color belong to skin and non-skin class. To compute the value,

Bayes rule is used:

19





s (c )

P(c | skin ) 

Ts (13)

n (c )

P(c | Non  skin) 

Tn (14)







where s(c) is the pixels count in the c- bin color skin class and n(c) is the pixel counts

in the c-bin color of non-skin class. It has been proved that certain separation exists

among skin and non-skin class. This leads to the building of fast and accurate skin

classifiers for images that collected in unconstrained imaging environment if the

training dataset is large enough. A skin classifier can be built from Bayes Maximum

Likelihood (ML) that implies







P(c | skin )



P(c | non  skin) (15)

1  P( skin)

 k

P( skin) (16)







where 0    1 is the threshold value that can be adjusted among true and false

positive value. The threshold value is determined from the ROC (Receiver Operating

Calculation) curve based on the training dataset. The ROC curve shows the

relationship between true and false values. The choices of prior probabilities do not

affect the detection behaviour as it is possible to choose an appropriate value of K

given same threshold value. This method is also applied in (Sigal et al, 2000).







Self Organizing Map is a type of neural network that is trained using

unsupervised learning to produce mapping that seeks to preserve the topological

properties of input spaces. These map units usually form a two-dimensional regular

lattice where the location of each unit carries semantic information. This advantage

allows SOM acts as a clustering tool of high dimensional data. The SOM is also in

favour because of its generalizing capabilities. Figure 2.6 shows the skin and non-

skin u-matrix produced from SOM method applied in Brown et al (2001).

20









Figure 2.7: The skin/non-skin u-matrix.







Brown et al (2001) used various colorspaces to compare the performance of

the two different SOM classifiers. The results show that the performance of SOM

classifiers does not change much corresponding to different colorspaces. At the

same time, it also showed that the choice of colorspace plays much important roles of

GMM models performance than SOM. It has been stressed that the SOM methods

need less resources than histogram and mixture models and is efficiently

implemented for run-time applications.









2.3.3 Parametric Skin-Color Modeling





Parametric skin color modeling is developed to adequate the need for more

compact skin model representation. Besides that, requirement for a model which has

the ability to generalize and interpolate the training leads to the development of this

model.

21





2.3.3.1 Single Gaussians







Under certain lightning condition, skin color distribution can be modeled

through elliptical Gaussian Joint probability density function (pdf), defined as







1 1

P (c )  1 1

 exp[  (c   ) T  1 (c   )]

2 2

2 (17)

2 |  |







where c is the color vector, µ and is the mean vector and diagonal covariance

matrix, respectively. The probability p(c) is used to calculate the skin-color

likelihood. The differentiation can be either obtained through the comparison of

threshold value estimated from training data or comparing Mahalanobis distance λ,

from the image pixel c to a certain threshold. This threshold can be judged from the

ROC curve calculated from training data where







  (c   ) T  1 (c   ) (18)









2.3.3.2 Mixture of Gaussian (MoG)







It is stated by Ruiz-del-Solar and Vershae (2004) that less training samples

are needed to obtain good probabilities estimator in MoG. It is considered to be a

more sophisticated model that capable to describe complex-shaped distributions. A

Gaussian Mixture density function is expressed as







N 1 1

P(c )  i 1 wi 1 1

 exp[  (c   ) T  1 (c   )]

2 2

2 (19)

2 |  |

22





where N is the number of mixture components and the weighted factor, wi is the



contribution of ith Gaussians. This model training is performed through using

iterative Expectation Maximization (EM) technique which ensures that the N-

component to be known beforehand. A critical part of this method is the selection of

component number N as different researchers have chose to use different values of N,

ranging from 2 to 16.









2.3.3.3 Elliptical Boundary Model





Elliptical boundary model is introduced as an alternative to solve the high

positive rate encountered in Symmetric Gaussian model. The performance of this

method is comparable to MoG and is as simple as SGM in computational complexity.

It is defined as







 (c)  [c   ]T  1 [c   ] (20)







where c is the color vector and  and  is the model parameter defined as





n

1 n 1

  ci ,   f (c i i   ) (c i   ) T

n i l N i 1 (21)







where N is the total number of samples in the training data set, fi is the number of



samples with chrominance ci while  is the mean of the chrominance vectors in the



training data set. The pixel with chrominance c is classified as skin pixel if  (c)   ,

where  is the threshold value chosen as trade off between true and false positives.

The disadvantage of this method is that it can only be used with binary classification.

This method has been applied in (Lee and Yoo, 2002).

23









Table 2.2: Comparison of Classifier

Classifier Category Paper Colorspace Advantages Disadvantages Justification

2.3.1Explicitly Defined Human Skin Colour Clustering For RGB, YCbCr 1. Simplicity of skin 1. Too many pixels are labeled 1. A robust decision rules

Skin Region Face Detection (Peer et al, 2003) detection rules under standard illumination will make this method a

when using YCbCr favorite to detect skin

colorspace under standard

Automatic Feature Construction And Normalized 1. Better performance 1. Do not consider the effect of illumination level

A Simple Rule Induction Algorithm RGB than SPM in rules with more than one 2. Threshold complexity

For Skin Detection (Gomez and precision and success selected criterion added when considering

Morales, 2002) rates factor added



Active Face And Feature Tracking HSV 1. Skin area is easily 1. Hue variance is small and

(Jordao et al, 1999) detected within fixed.

threshold 2. Skin color of different races

may added the threshold

complexity

Nonparametric Skin Estimation And Prediction OF HSV 1. Stable to rapid 1. The detection performance is 1. Good to use for detection

Distribution Modeling Evolving Color Distribution For Skin changes in the low if the algorithm is over in changing background as

Segmentation Under Varying appearance of skin segmented the effect is less compare

Illumination (Sigal et al, 2000) color due to with foreground

illuminant changes, 2. Fast in training and usage

surface inflection and 3. Much storage space are

rapid changes in required

background 4. Unable to generalize

training data

Statistical Color Model With RGB 1. Better performance 1. Need lots of training data in

Application To Skin Detection (Jones than Gaussian model order to achieve better

and Reg, 1999) in speed and performance

accuracy

2. Demonstrate

separability between

skin and non-skin

pixel

A SOM Based Approach To Skin HSV, 1. Less dependant on 1. Face difficulty when dealing

Detection With Application In Real Normalized the choice of with non-skin samples

Time System (Brown et al, 2001) RGB configuration than

mixture model





23

24









Classifier Category Paper Colorspace Advantages Disadvantages Justification

Parametric Skin Face Detection On Color Image (Hsu YCbCr 1. Able to detect dark 1. Performance decrease when 1. The detection performance

Distribution Modeling et al, 2002) and light skin tone detecting group of people in varies significantly from

color an image colorspace to colorspace

2. Able to detect skin color

distribution under fixed

Comparative Performance of Different TSL, 1. Performed well with 1. Problem in detecting dark lightning condition

Chrominance Spaces For Color Normalized the TSL colorspace skin in normalized

Segmentation and Detection of Human RGB 2. Successfully colorspace

discriminate face 2. False positive occur due to

Faces in Complex Scene Image

cluster limited color discrimination

(Terrillon et al, 2000) of SGI camera



An Elliptical Boundary Of Skin Normalized RG 1. Better performance in 1. Usage is limited to binary

Detection (Lee and Yoo, 2002) accuracy than Single classification as the

Gaussian and continuous information

Mixture of Gaussian given by probability density

function is not given









24

25





2.4 Fuzzy Methods







Besides explicitly defined skin region, parametric and non-parametric method,

fuzzy methods also emerge as popular techniques in skin detection. Many fuzzy

approaches have been proposed to perform skin detection. In this project, we are

going to focus on fuzzy approaches which use color as features.









2.4.1 Takagi Sugeno Fuzzy Inference System







Hamid and Jemma (2006) proposed a Takagi Sugeno fuzzy inference system

(FIS) for pixel classification. This system consists of two inputs which are Cb and Cr

and one output which is either skin or non-skin. Each input is categorized into three

sub-sets, light, medium and dark. Then, If-Then rules are applied in each pixel in the

image in order to decide whether the pixel belongs to skin or non-skin region. The

fuzzy rules applied are







i. If Cb is Light and Cr is Light, then the pixel = 0.

ii. If Cb is Light and Cr is Medium, then the pixel = 0.

iii. If Cb is Light and Cr is Dark, then the pixel = 0.

iv. If Cb is Medium and Cr is Light, then the pixel = 0.

v. If Cb is Medium and Cr is Medium, then the pixel = 1.

vi. If Cb is Medium and Cr is Dark, then the pixel = 1.

vii. If Cb is Dark and Cr is Light, then the pixel = 0.

viii. If Cb is Dark and Cr is Medium, then the pixel = 1.

ix. If Cb is Dark and Cr is Dark, then the pixel = 0.







After that, determine the degree of membership to appropriate fuzzy sets

through membership functions. Once the inputs have been fuzzified, the final

decision of the inference system is the average of the output ( z i ) corresponding to

26





the rule ( ri ) weighted by the normalized degree pi of the rule. Figure 2.7 shows the



input and output of the images.









(a) (b)

Figure 2.8: Takagi Sugeno Fuzzy Inference System (FIS) (a) Input image (b) Output

image









2.4.2 Modified Fuzzy C-Mean Algorithm (MFCM)





A modified fuzzy clustering method was proposed by Chahir and Elmoataz

(2006) to classify pixels within an image into C class by minimizing the objective

function below.





C N C

J   (u ij ) 2 d 2 ( x j , ci )    pi log( p i )

i 1 j 1 i 1 (22)





N

1 Ni

where Pi 

N

u

j 1

ij 

N

. N is the cardinality of cluster i, Pi is interpreted as the



2

probability of cluster i. d ij  d 2 ( x j , ci ) is the standard Euclidean Distance between



ith centroid ( ci ) and jth pixel.







In order to take account of distribution of Hue, Saturation, Value and spectral

distribution of pixel, D, distance function was chosen as

4 4

N ik

d ij   wk ( x jk , cik ) and Pi   . wk serves as weighting constants between the

k 1 k 1 N

27





different channels. cik and x jk represents centroid of class i in channel k and jth pixel



in the same channel. N ik is interpreted as the cardinality of cluster I in channel k.







C

1 1

2

d ij 

 (d

k

2

log Pk )

kj

u ij  C

 2

[log Pi  C

]

1 2 Nd ij 1 (23)



k

( 2)

d kj

k

( 2)

d kj







After updating centroids and membership function, update the prior

probabilities of clusters. Discard the cluster whose probabilities are given threshold 

which must be small enough to discard non-significant clusters. Compute the

Euclidean distance, d ij between centroids and pixels. If | J n  J n1 |  , then

2







n  n  1 , update  and membership matrix (U)









2.4.3 Fuzzy Fusion







Fuzzy Integral was introduced by Sugeno which said to be outperforms other

fusion operators such as maximum, weighted sum or order weighted averaging. The

outperformance resulted from the ability of fuzzy integral to compute a large

diversity of results. The results from the increment number of parameters have to be

tuned to produce the outperformance. As a result, the relevance of methodologies for

parameterizing the operator in an automated manner increases.





There are two popular approaches in fuzzy integral, the Sugeno Fuzzy Integral

and the Choquet Fuzzy Integral. The Sugeno Fuzzy Integral is defined as





n

S  [ X 1 , , X n ]  i 1[h(i ) ( xi )   ( A( i ) )] (24)

28





While the Choquet Integral is defined as





n

C  [ X 1 , , X n ]   h(i ) X i [  ( A(i ) )   ( A(i 1) )]

i 1 (25)







where  ( A(0) )  0 . The xi i  1,2,3 denotes the image gray level of each color



channel in a particular color representation, h(i ) denotes the fuzzification function



applied on these data and  ( A(i ) ) is the weight of the importance of color channel in



the skin detection.







 ( A j )   ({ X 1 , , X n }) j  1, , n (26)







A fuzzy measure  presents 2 n coefficients  ( A j ) j  1,  ,2 n as subset



A j was formed among the input information sources. From all coefficient, just n are



taken into account in each fuzzy integration  ( A j )   ({ X 1 , , X n }) j  1, , n



which are selected upon the sorting operation denoted by the enclosed subindices.







In fuzzy integral, each canonical region of features space is defined a set of

weighting coefficient to improve the robustness of the operator. The robustness of

the fuzzy integral indicates the facts that whenever there is a change in the

illumination condition, it modifies the grayvalue of the pixels in the color channels

instead of the ordinal relationship among these grayvalue. Fuzzy fusion has been

applied by Soria-Frisch et al (2007) as an approach to improve the result of different

methodology taken by its own. This performance of this method has been

outstanding compared with statistical model in (Hsu et al, 2002) and (Jones and Rehg,

1999).

29









Figure 2.9 Block diagram of the framework for the detection of skin based on the

fusion paradigm









2.4.4 Linear Matrix Inequality (LMI) Fuzzy Clustering







M. H. Kim et al (2005) had proposed to use LMI fuzzy clustering to

determine the rule structure and parameters of the membership function which based

on Takagi-Sugeno fuzzy classifier. In the proposed method, the number of cluster

becomes the number of fuzzy rule and the membership functions in each rule are

treated as fuzzy cluster function. As a result, the firing strength of output is

represented by the clustering method instead of fuzzy set.







Consider a finite set of input features X  {x 1 , x 2 , , x N d } , x k  F as



classification data. x i is the ith input feature set containing HSI pixel information.

N c is the given number of fuzzy cluster. The result of LMI clustering is expressed

by a partition U as







U  [u ij ], i  1,  , N c , j  1,  , N d (27)







where u ij is a crisp numerical value in [0,1] and express the membership of element



x j belongs to the ith cluster. Partition matrix has two constraints which are

30





Nc



u ij  1 , for all j  1, 2,  , N d

i 1 (28)

Nd

0   u ij  N d ，for all i  1,2,  , N c

j 1 (29)







In addition, a cluster additional membership function matrix has been used to

determine data partition. The membership matrix W is defined as







W  [ wij ], i  1, , N c , j  1, , N d (30)







where wij is the firing strength of element x j belongs to the ith cluster.





i

( mk  x kj ) 2

n  i

k

wij   e (31)

k 1









After membership matrix is calculated, the partition matrix is updated by using

membership degree







u ik  1, i  1,2, , N c (32)

Nc

k  arg j , j  C i , max  u hj  0

h 1 (33)







where C i is the set of data belongs to class of the ith cluster.





When partition matrix is updated, the width and the center of membership

i

function are updated via LMI optimization method. The center mk is simply updated



by averaging x kj belongs to the ith cluster. The width update can be denoted as



identifying membership function satisfying following two constraints.

31





i  1, x k  C i (34)



i  0, x k  C i (35)







For computational convenience, the parameters of i ( x k ) can be reformulated as



follows





i

i

( c1  x k ) 2

z

i i

( c2  x2 ) ( c3  xn )

 i

 i

 i

v1 v2 vn

i ( x k )  e e xe

3 ( c ij  xik ) 2

  v ij (36)

e j 1





k

 ci )T Vi ( x k  ci )

 e ( x

k T

) Vi ( e k )

 e ( e







1 1

where Vi  diag ( i

, , i ) is the diagonal matrix containing the widths of the

V1 V3



Gaussian membership functions of the ith cluster, and ci  [c1i ,  , c3 ] represents

i







center values of the membership function of the ith cluster.

32









Table 2.3: Comparison of Fuzzy Methods on Skin Detection

Method Paper Colorspace Techniques Advantages Disadvantages Justification

2.4.1 Takagi Fuzzy Classification, YCbCr i. Three input membership iii. Overcome the i. Insufficient to detect i. Simple method to

Sugeno Fuzzy Image Segmentation function is defined – light, limitation of skin in different apply

Inference and Shape Analysis medium and dark classic method illuminance condition ii. Insufficient to classify

System for Human Face ii. The skin range for Cb and which face is ii. Not sufficient to be used skin and non-skin

Detection (Hamid and Cr based on the image correctly identified to detect face pixel in different

Jemma, 2006) database fall within illumination

[77,127] and [139, 210]

2.4.2 Skin-color Detection HSV i. Fuzzy c-mean is used to iv. A more compact i. Time consuming as time i. Achieve optimum

Modified using Fuzzy update the centroid of a and more is needed to generate a performance if a

Fuzzy C- Clustering (Chahir skin cluster homogeneous suitable rule to perform robust cluster is

Mean and Elmoataz, 2005) ii. Update skin probability of cluster can be skin detection identified

Algorithm each cluster determined ii. Time consuming

(MFCM) iii. Calculate the Euclidean when size of dataset

distance between centroid increase

and pixel which high

membership value is

assigned to color value

which close to centroid



2.4.3 Fuzzy Fuzzy Fusion for Skin RGB i. A new parameter is i. Outperform i. Falsely detect some i. Integration of two

Fusion Detection (Soria- denoted to the fuzzy method proposed background pixel as different methods

Frisch, 2007) measure in order to by Jones and Rehg skin pixel allow it posses the

integrate MoG, Hsu [5] and Hsu et al advantages of both

methodology and DoFI [11] methods

ii. Genetic algorithm is ii. Performance is

applied to assess the fuzzy unknown under

measure coefficient different illumination

level.

2.4.4 Linear New Skin Model for HSI i. Gaussian membership i. Different skin i. Time consuming in i. Performance on

Face Detection (Kim function is used clusters correspond order to generate different dataset is

Matrix

et al, 2005) ii. Center and width of to different superior rule and unknown.

Inequality membership function is illuminance are membership function

updated by LMI Fuzzy identified

(LMI) Fuzzy

Clustering

Clustering







32

33





2.5 Summary







It has been proven that every color space has an optimum skin detector which

the performance of all these skin detectors is the same. Therefore, the performance of

the skin detection scheme is actually depending on the design of the detection

scheme.





Different skin detection schemes yield different performance as each

possesses different advantages and limitation. The common limitations of skin

detection schemes are the existence of pixels with skin-like color and the changing

illumination condition. The fuzzy fusion is designed to combine different detection

methodologies as each provides different pieces of evidence to solve the limitations.

In this project, fuzzy logic will be used to classify pixel into skin and non-skin pixels.

Finally, the performance of the method proposed in this project will be compared

with Explicitly Defined Skin Region Model by Peer et al (2003) and Takagi Sugeno

Fuzzy Inference System by Hamid and Jemma (2006).

CHAPTER 3









METHODOLOGY









3.1 Introduction







In this chapter, the methodology that will be proposed in this project will be

discussed. Methodology is a systematic analysis of methods and procedures that will

be applied to achieve the objectives stated in Chapter 1. Through the study of

methodology, the idea of how this project is implemented can be identified. In Figure

3.1, we show the process of skin detection. The proposed methodology used in this

project will be discussed in the section 3.2. Section 3.3 will discuss result evaluation

method. Then, the conclusion for the proposed methodology will be discussed in

section 3.4.

35







Color Features

Preprocessing Extracted





Web Pages



Image

Features

Selection









Color Features







Classification

Result Classified By

Subtracted Image







Fuzzy Classifier





Figure 3.1: General skin detection process









3.2 Proposed Methodology







As mentioned in Chapter 1, this project is focus on solving the problems that

skin detection encountered when using color as cue. One of the problems faced

during skin detection is the color likelihood problem between background color and

skin color. This problem occurs when a pixel contains similar color information as

human skin color information which causes the computer hard to establish the

differences between them. Therefore in this chapter, a methodology was proposed to

solve this problem. This methodology consists of five modules which are data

collection, preprocessing, feature extraction, image segmentation and classification.

These five modules are shown in Figure 3.2 and how these five modules explore the

objectives stated in Chapter 1 is illustrated in Table 3.1. Further detail of these

modules will be explained at section 3.2.1, 3.2.2 and 3.2.3.

36







Part a: Data Part c: Feature Extraction

Retrieval Colorspace Selection



Web Pages



Part b: Skin Modeling

Preprocessing Processed Image









Color Features

Color Region

Part e: Fuzzy

Classification Part d: Background Subtraction









Animal Image Human Skin

Group Image Group







Figure 3.2: Methodology of this project







According to Figure 3.2 shown above, the following steps will be taken in

this methodology:







a. Retrieve data using web-crawler.

b. Resize the images to smaller dimension.

c. Extract features from the images.

i. Determine suitable colorspace

ii. Build skin color model

d. Subtract background in order to get the object of interest

e. Classify whether an image belong to human skin or animal image based on

fuzzy classification







The methodology proposed starts from web pages retrieval using web-crawler.

The retrieved web pages are then preprocessed to filter undesirable noises. After

preprocessing, text will be filtered and only images will be left. These images then

were used for feature extraction. To do feature extraction, the suitable colorspace

37





must first be determined. Then, the color value is extracted from the images and the

skin model is built by using color histogram method.







Once features extraction completed, the image is subtracted to isolate the

object of interest. Finally, the subtracted images will be classified by fuzzy classier to

determine whether pixels within images belong to skin pixel.





Table 3.1 Modules in methodology

Modules Objective Problem Process

Statement

a) Data Retrieval - - - Retrieve web

pages that

contain images

b) Preprocessing - - - Filter

undesirable

noises and text

which left only

images

c)Feature Extraction - - - Determine the

i. Colorspace selection suitable

ii. Skin Modelling colorspace to be

used in this

project.

- Extract color

features using

color histogram.

d) Background - - - Subtract image

Subtraction to isolate the

object of

interest

e) Fuzzy Classification i. To study human and - How fuzzy classifier - Classify

animal skin images using able to discriminate whether an

fuzzy classification human and animal skin image belongs

ii. To classify the images to human skin

detection performance - How fuzzy classifier or animal

when dueling with porn differentiate porn and image group

and non-porn image non-porn images

38





3.2.1 Data collection







The skin detection process starts from web page retrieval. These web pages

are retrieved by using web crawler. The web-crawler to be used in this project is

Webripper which is an open-source. The web crawler will be used to periodically

search and download web pages, adult sexual images and suspected images from the

internet. After all the suspected web pages have been attained, the attained data will

be sent to preprocessing stage to eliminate unwanted noise.







Table 3.2 Explanation of data collection

Process Explanation Techniques

Data retrieval Web pages that contain images will be retrieved from i. Web Ripper

various sources. At the same time, suspected images

will also be downloaded.









3.2.2 Preprocessing







The downloaded web pages will be filtered to get rid of text and only images

will be left. Then, images other than Asian ethnic are filtered as well as this project

will only focus on Asian ethnic within this project. After that, the images that contain

human and animal skin will be cropped into images with pure human skin and

animal skin. These images will later be resized into 24x24 dimensions. The

preprocessing steps are explained below. Table 3.2 gives a brief explanation of these

two modules.







Step 1 : Filter downloaded web pages to get rid of text.

Step 2 : Select images with Asian ethnic only.

Step 3 : Cropped selected images into images with pure human and animal skin

Step 4 : Resize each cropped image into 24x24 dimension

39





Table 3.3 Explanation of preprocessing

Process Explanation

Preprocessing Web pages downloaded are filtered to get rid of undesirable noises. The

undesirable noises include text, html tags and thumbnails. The images will

then be scaled to lower dimension.









3.2.3 Feature Extraction







After the preprocessing step, we need to extract the features from the images.

The features that we intend to extract are color information. The color information

extracted is important to use for classification later. To extract color information, we

must first determine suitable colorspace. Table 3.3 will give a brief description of

feature extraction









3.2.3.1 Colorspace Selection







There are many existing colorspaces which are used to represent color in

either video or images. Some of the popular colorspaces are RGB, HSV, YCbCr and

CIE-Lab. The selection of colorspace will affect the performance of the classification

in the later stage. In this project, the selected colorspace is RGB which is popular

because of the fact that it is easy to use. The RGB value corresponds to the red, green

and blue information of the particular pixel within images.









3.2.3.2 Skin Modelling







After the colorspace selection has been made, the color features will be

extracted from the images and a skin color model will be built. The purpose of

building skin color model is to identify the skin color distribution. Since RGB

40





colorspace is used to extract color features, the value corresponds to the three color

channels which are red, green and blue will be acquired.

The modelling method used in this project is color histogram. It was proposed

by Jones and Reg (1999) as a statistical model to detect skin and non-skin pixels. It is

a standard statistical description of the frequencies of color occurrence within an

image. To create a color histogram for human skin, build a table with 256x256x256

dimensions. These three columns are corresponding to red, green and blue value

while 256 is the range of RGB value from 0 to 255. After that, get the RGB of each

pixel and add up 1 to the corresponding color channel in the table. Then, calculate

the skin discrete probability, P(c) of each color channel to create the color histogram.







skin(c ) (37)

P (c ) 

Norm







where skin(c) is the total count of each color channel and Norm is the maximum

count of all color channel .The skin modelling steps are explained below.







Step 1 : Build a table with dimension of 256x256x256 which corresponds to red,

green and blue color value.

Step 2 : Get the RGB value of each pixel

Step 3 : Fill the table by adding value of 1 to corresponding color channel

regarding to RGB value of each pixel

Step 4 : Calculate the skin discrete probability, P(c) as in (37)

Step 4 : Build color histogram base on P(c) of each color channel







Table 3.4 Explanation of feature extraction

Step Process Explanation

i Colorspace Selection Select suitable colorspace from existing colorspace. The

performance of colorspace in classification will be evaluated.

ii Skin Modelling Calculate the skin discrete probability of each color channel to

identify the skin color distribution.

41





3.2.3.2.1 Integration of Skin Modeling in Fuzzy Rules





The color histogram for each color channel was built by calculating skin

discrete probability of each channel within RGB value. Based on (37), the highest

skin discrete probability within the color channel of RGB value can be determined.

The color channel with highest skin discrete probability of each color histogram was

defined as the center of Gaussian membership function. Then, by justifying the

distance between the color channels with skin discrete probability more than 0.3, the

width of the Gaussian membership function is identified. In the example given in

Figure 3.3, the highest skin discrete probability of red color channel fall in 231. So,

value 231 is defined as the center of Gaussian membership function. Meanwhile, the

range of skin distribution is within 160 to 250. Therefore, the width of the Gaussian

membership function is 250 -169 which is 81.









The Highest Skin Discrete

Probability within color

histogram is defined as

center, c = 231









The range of skin distribution with skin

discrete probability => 0.3, is defined as

width,



a = 250 – 169



= 81





Figure 3.3: Integration of skin modeling in fuzzy membership function

42





3.2.4 Background Subtraction







Background subtraction technique has been adopted in many vision-based

interfaces to extract or track moving objects. This technique intends to separate the

user of interest from the image in order to convey user intention properly. In this

project, the background subtraction will be implemented based on the technique

stated by Hong and Woo (2003).







To implement the background subtraction, we must first get the RGB value of

each pixel within the image. Then, we get the mean value and standard deviation of

each pixel. The following equations show how the vector of mean and standard

deviation of each pixel is calculated







1 N 1

 i ( R, G , B ) 

N

 0

I i ( R, G , B )

(38)



1 N 1

 i ( R, G , B )   0

( I i ( R, G , B)   i ( R, G , B)) 2

N (39)







where I i ( R, G , B) represents the vector of pixel i’s color channel in RGB and N is

the number of pixels within an image. With the mean vector and standard deviation

of each pixel color channel calculated, the threshold value is determined in order to

subtract the background scene from the object of user interest. The threshold value of

pixel i is mapped by function of standard deviation of pixel i which are shown in (40).







Thi ( R, G, B)     i ( R, G, B) (40)







where Thi is the threshold value of pixel i in RGB while is the determinant



constant which determines the confidence interval. For this project,   2 as it has

about 95% of confidence interval that decides the threshold ranges.

43





To compare the differences between color channel and threshold value of

pixel i, equation (41) is applied which is shown below.



Fi  u ( Di ( R)  Thi ( R))  u ( Di (G )  Thi (G ))  u ( Di ( B)  Thi ( B)) (41)





where Fi is the determinant function which characterizes pixel i in each colorspace.

At the same time, u is a unit step function which value is either 0 or 1. If

Di ( x )  Thi ( x) , then u is 1. Otherwise, it is 0.







The pixel will then be characterized into two groups based on the equation

(42) as follow.





B : 0  Fi  c1

Ii   s (42)

H : Fi  c1





where B is the background image and H s is the segmented user images with shadows.

c1 is the determinant constant which is set as 2 in this project to subtract the

background from the object of interest. The background subtraction steps are

explained below. Table 3.4 explains the purpose of applying background subtraction.







Step 1 : Get the RGB value of each pixel within image.



Step 2 : Calculate the mean vector and standard deviation of each pixel as in (38)

and (39).



Step 3 : Determine the threshold value, Thi by mapping standard deviation and

selected determinant constant,  .



Step 4 : Calculate the determinant function, Fi as in (41) and determine u.



If Di ( x )  Thi ( x) , then u = 1 else u = 0.



Step 5 : Characterize pixel i.

If 0  Fi  c , I i is background image, B.



If Fi  c , I i is segmented user images, H s

44





Table 3.5 Explanation of Background Subtraction

Process Purpose Proposed

Techniques

Background Separate the object of interest from background i. Background

Subtraction for further image processing Subtraction for

Vision-Based User

Interface









3.2.4 Fuzzy Classification







This is the core part of this project which a fuzzy classifier was proposed to

solve the likelihood problem of skin detection. This classification technique will then

be used to determine whether an image belong to human or animal image. The main

reason of using fuzzy classifier is to solve the problems that stated earlier in Chapter

1. Considering the color representation in different illumination levels and likelihood

between background and skin color add complexity in defining the boundary

between skin and non-skin pixels in term of color. This complexity leads to the

question of fuzziness. Therefore, fuzzy theory appears to be a good solution in

solving the fuzziness in skin detection problem.







The fuzzy theory was proposed by Lofti Zadeh in 1965. The implementation

of fuzzy theory starts with defining fuzzy sets. In this project, three inputs were

defined which are I Re d , I Green and I Blue which correspond to the RGB colorspace.



Then, the membership degree of each input is identified by applying membership

function.







Membership function is a curve that defines how each point of input is

mapped to a membership value between 0 and 1. The function can be an arbitrary

shape whose shape suits to different problems from the prospect of simplicity,

convenience, speed and efficiency. As Gaussian membership function had been used

in (Tasaltin and Taylan, 2007) and (M. H. Kim et al, 2005) to determine the human

skin membership degree of color value, the Gaussian membership function will be

used in this project to define the membership value. The Gaussian membership

45





function is popular for specifying fuzzy set because of their smoothness and concise

notation. The membership degree of each input I ( R ,G , B ) belongs to fuzzy set High,



FHigh is,



1

 High  (43)

I ( R ,G , B )  c 2b

1 | |

a

where a and c is the width and center of fuzzy set High, FHigh . Meanwhile, the



membership degree of input I ( R ,G , B ) to fuzzy set Low, FLow can be find based on



equation (34).

(c  I ( R , G , B ) ) 2

 Low ( I ( R ,G , B ) )  exp(  ) (44)

2 2



where c and  is the center and width of fuzzy set . The width and center value of

each membership function is determined from the skin modeling. After specifying

the membership values of I ( R ,G , B ) to FHigh and FLow , the fuzzy rules are interpreted. To



do so, first fuzzifies the input. Then, applies the operator to the antecedents and

shapes the fuzzy set using the degree of support. The fuzzy rules are set as below.





Rules 1 : If I Re d is FHigh and I Green is FHigh and I Blue is FHigh , then ORules1 is FSkin



Rules 2 : If I Re d is FHigh and I Green is FHigh and I Blue is FLow , then ORules 2 is FMedium



Rules 3 : If I Re d is FHigh and I Green is FLow and I Blue is FHigh , then ORules 3 is FMedium



Rules 4 : If I Re d is FHigh and I Green is FLow and I Blue is FLow , then ORules 4 is FNon  Skin



Rules 5 : If I Re d is FLow and I Green is FLow and I Blue is FLow , then ORules 5 is FNon  Skin



Rules 6 : If I Re d is FLow and I Green is FHigh and I Blue is FHigh , then ORules 6 is FMedium



Rules 7 : If I Re d is FLow and I Green is FHigh and I Blue is FLow , then ORules 7 is FNon  Skin



Rules 8 : If I Re d is FLow and I Green is FLow and I Blue is FHigh , then ORules 8 is FNon  Skin





where ORulesi is the output of each rules which determine the output fuzzy set belongs



to fuzzy set Skin, FSkin , Near-to Skin, FMedium or Non-skin, FNon  Skin . The output fuzzy

46





set will then be aggregated to a single output fuzzy set. The aggregation of rules is as

follow.







 Skin  ORules1 (45)



 Medium  ORules 2  O Rules 3  ORules 6 (46)



 Non  Skin  O Rules 4  ORules 5  ORules 7  O Rules8 (47)







Finally, the resulting set is defuzzified to get the number. The defuzzification

method used in this project is center of gravity (COG) which is stated in (48).





b





a

A ( x ) xd ( x )

COG  b

(48)

 a

A ( x)d ( x)









The number will then determine which group that the image belongs to. The

output number is divided into three groups which are Skin, Near-to Skin and Non-

skin. The output number within 0 to 0.39 is described as Non-skin pixel while output

number within 0.4 to 0.69 is defined as Near-to Skin. The output number of the last

group which is the Skin pixel falls within 0.7 to 1.0. The steps of fuzzy classification

are explained as below.







Step 1 : Three inputs are defined which are I Re d , I Green and I Blue correspond to



RGB value.

Step 2 : Determine the membership degree of each input,  High ( I ( R ,G , B ) ) to



fuzzy sets High, FHigh and the membership degree,  Low ( I ( R ,G , B ) ) to



fuzzy set Low, FLow .



Step 3 : Three output fuzzy sets are identified, which are Skin, FSkin , Near-to



Skin, FMedium and Non-skin, FNon  Skin .

47





Step 4 : The if-then rules are set up using logical operator. The set up rules are

i. O Rules1   High ( I Re d )   High ( I Green )   High ( I Blue )



ii. O Rules 2   High ( I Re d )   High ( I Green )   Low ( I Blue )



iii. O Rules3   High ( I Re d )   Low ( I Green )   High ( I Blue )



iv. O Rules 4   High ( I Re d )   Low ( I Green )   High ( I Blue )



v. O Rules5   Low ( I Re d )   Low ( I Green )   Low ( I Blue )



vi. O Rules 6   Low ( I Re d )   High ( I Green )   High ( I Blue )



vii. O Rules 7   Low ( I Re d )   High ( I Green )   Low ( I Blue )



viii. O Rules8   Low ( I Re d )   Low ( I Green )   High ( I Blue )



Step 5 : Aggregate all the output using (45), (46) and (47) which the output of

each rule is combined into single fuzzy set.

Step 6 : Defuzzification is implemented based on (48) to convert the fuzzy

output into a crisp number.







Table 3.6 Explanation of Fuzzy Classification

Process Explanation Techniques



Fuzzy Classification Classify an pixel within image either i. Fuzzy Logic

belongs to skin or non-skin pixels









3.3 Result Evaluation Method







A sample is classified as true positive if both the predicted and actual value is

true. It is defined as false positive if the predicted value is positive but the actual

value is negative. Conversely, if it is predicted to be negative and the actual value is

positive, it is defined as false negative. However, if the predicted and actual value are

both negative, it is classified as true negative. The result is later represented in

Receiver Operating Characteristic curve (ROC) which T. Fawcett (2007) gave a

detail analysis.

48





The ROC curve allows us to analysis and selects the optimal model that best

fits the method. To draw ROC curve, the true positive rate (TPR) and false positive

rate (FPR) need to be identified. TPR which is also referred as sensitivity measures

the classifier performance to check the percentage of correctly classified positive

images among all positive samples. TPR is defined as







TP

TPR (49)

(TP  FN )







where TP is the total amount of true positive samples and FN is the total amount of

false negative samples. On the other hand, FPR calculates how many incorrect

positive results occur among all negative samples. It is defined as







FP

FPR 

( FP  TN ) (50)







where FP is the total amount of false positive samples and TN is the total amount of

true negative samples.





The Area under ROC curve can be found by constructing a trapezoid under

the curve. The Area under ROC is used to measure the discrimination, which is the

ability to classify whether an image is fall into skin image group or non-skin image

group. T. G. Tape has stated a traditional point system that measure the Area under

ROC curve which is shown as below.







i. 0.9 – 1 = excellent

ii. 0.8 – 0.9 = good

iii. 0.7 – 0.8 = fair

iv. 0.6 – 0.7 = poor

v. 0.5 – 0.6 = fail

49





Meanwhile, the positive likelihood ratio and negative likelihood ratio are

evaluated when the highest accuracy is achieved. Positive likelihood ratio is the

probability of a positive test result given the presence of the human skin pixel and the

probability of a negative test result in the absence of human skin pixel. Negative

likelihood ratio is the probability of a negative test result given the presence of the

human skin pixel and the probability of a negative test result in the absence of human

skin pixel.







True Positive Rate

Positive Likelihood Ratio 

False Positive Rate (51)

False Negative Rate

Negative Likelihood Ratio 

True Negative Rate (52)





Then, the P-value is obtained from the experiment which is the probability of

an observed Area under ROC curve is found. A classification method with a P- value

lower than the true Area under ROC curve is considered good as it is able to

discriminate two different groups of images.









3.4 Summary







This chapter proposed the methodology that will be implemented in this

project. This methodology includes five main modules which are data collection,

preprocessing, feature extraction, image segmentation and fuzzy classification.

Section 3.2 shows the general view of the methodology. Section 3.2.1 explains how

the data is collected and how the collected data is preprocessed. Meanwhile, Section

3.2.2 discussed about the feature extraction method used in this project. Then, the

detail of image segmentation is stated in Section 3.2.3. The last module which is the

fuzzy classification is further explained in Section 3.2.4. Then, the result evaluation

method will be discussed in Section 3.3. Finally, this project’s objectives are

expected to be achieved through the combination of these five modules.

CHAPTER 4









RESULT AND DISCUSSION









4.1 Introduction







In this chapter, the experiment implemented in this project will be described.

Then, the experiment result will be analysed. The main focus of the experiment is to

identify how the fuzzy rules conducted will affect the performance of the skin

detection. Section 4.2 will give a brief description of how the experiment is

implemented while Section 4.3 illustrated the steps taken to implement the

experiment. At the same time, Section 4.4 shows the result of the experiment while

the result comparison is shown in Section 4.5. Then, Section 4.6 discussed the result

evaluated. Finally, the conclusion for this chapter is shown in Section 4.7.









4.2 Skin Detection in Images







The experiment conducted is based on the methodology stated in chapter 3

which includes three main parts. These three main parts are feature extraction,

background substraction and fuzzy classification. The feature extraction is the first

process to be run through which intends to extract color features from the input

images. These extracted features will later be used in fuzzy classification to identify

the skin pixels within the image. At the same time, skin color boundary is identified

by building the skin model using RGB colorspace and color histogram. The skin

color boundary is associated to the setting of fuzzy rules which plays an important

part in fuzzy classification. Then, the image background is subtracted in order to get

51





the relevant object within the image. After the image is subtracted, only the

remaining object will go through fuzzy classification. The classification result of

each pixel will then be grouped up to decide whether human exists within an image.

There are four main processes in this experiment which are:







i. Color Feature Extraction

ii. Skin Modelling

iii. Background Subtraction

iv. Fuzzy Classification





The framework applied to conduct this experiment is shown below





Color Feature Skin Modelling

Images

Extraction

Color

Feature

Skin Color

Boundary





Background Subtraction





Subtracted Image





Fuzzy Classification









Skin Pixels Total Up









Image Group







Figure 4.1: Experiment method

52





4.3 Experimental Setup







To setup the experiment, several approaches and parameters have been used.

Section 4.3.1 discussed the skin modelling while Section 4.3.2 shows how the

background subtraction is done. Then, Section 4.3.3 shows the implementation of

fuzzy theory in skin detection.









4.3.1 Skin Modelling







The range of skin color will be identified to be used as reference in fuzzy

classification to determine whether a particular pixel belongs to skin pixel. The range

of red, green and blue color channels is 0 to 255. The identification of skin color

range is done through the building of color histogram. The color histogram is built by

using statistical analysis of 100 pure human skin images. Figure 4.2 shows the skin

color distribution in Red channel while Figure 4.3 and Figure 4.4 show the skin color

distribution of Green and Blue channel. Table 4.1 shows the skin color boundary of

red, green and blue value.









Figure 4.2: Skin discrete probability of red color channel

53









Figure 4.3: Skin discrete probability of green color channel









Figure 4.4: Skin discrete probability of blue color channel







Table 4.1 Skin Color Boundary of RGB colorspace

Red Green Blue

Skin Color Boundary 169 – 250 140 – 210 50 – 208

Non-skin Color 0 – 168 & 251-255 0 – 139 & 211 – 255 0 – 49 & 209 – 255

Boundary

54





4.3.2 Background Subtraction







In this process, the background of the image will be subtracted in order to

identify the object of interest. Once the objects of interest are identified, the

subtracted image will proceed for classification in later stage. The background

subtraction technique used in this project will base on the technique that proposed by

Hong and Woo (2003). Figure 4.5(a) shows the image before background subtraction

while Figure 4.5(b) shows the image after subtraction. After background subtraction,

only human object remained while other parts of the image are blackened.









(a) (b)

Figure 4.5: (a) Image before background subtraction. (b) Image after background

subtraction









4.3.3 Fuzzy Classification







The input for fuzzy classifier is the pixel values correspond to the RGB

colorspace. The output of the classification will be the result of whether the

particular pixel is skin or not skin. This fuzzy classification is based on the Mamdani

fuzzy inference method (Anderson and Hall, 1999). The testing is conducted by

using vb.net as programming language and based on the Matlab Fuzzy Logic

Toolbox. The Mamdani inference method is shown in Figure 4.6.

55









Fuzzify Inputs









Apply Fuzzy Operator









Apply Implication Method









Aggregate Outputs









Defuzzification









Final Output







Figure 4.6: Mamdani fuzzy inference method









4.4 Experimental Result







The classification that will be done in this project is to classify the pure

human skin images and animal images. Then, classification is also done on porn and

non-porn images to test the usability of fuzzy classification in filtering unhealthy

web content. The result was obtained through fuzzy classification which tested on

four types of images stated above. Each type of images consists of 100 samples.

Therefore in this section, we will analysis the result based on the discriminate ability

of fuzzy classification. The analysis of experiment result of first classification will be

shown in Section 4.4.1 and the second classification in Section 4.4.2. Further result

discussion will be shown in Section 4.6.

56





4.4.1 Analysis of Fuzzy Logic Ability to Discriminate Human and Animal Skin

Images







In this analysis, we will evaluate the true positive rate (TPR) and false

positive rate (FPR) of classification result which is shown in Table C.1 as listed in

Appendix C. To determine the discriminate ability of fuzzy classification, we build

the Receiver Operating Characteristic curve (ROC) based on the TPR and FPR that

displayed in Figure 4.7.









100







80





Sensitivity: 68.0

True Positive Rate









Specificity: 93.0

60 Criterion : >80.85









40







20







0

0 20 40 60 80 100

False Positive Rate





Figure 4.7: Receivers Operating Characteristic Curve (ROC) of skin percentage of

human and animal skin images samples







Figure 4.7 shows that the Area under ROC curve (AUC) is 0.807 which can

be considered well based on the traditional academic point system. Although it did

not achieve the perfection, the ROC curve showed that the applied fuzzy rules are

able to discriminate skin and non-skin pixels. The highest accuracy of classification

57





achieved when the sensitivity is 68% and the specificity is 93% at criterion 80.85%

which is closest to optimal result at point (0,100).







With sensitivity equals to 68% and the specificity as 93%, the positive

likelihood ratio is 9.71 and the negative likelihood ratio is 0.34. As the positive

likelihood ratio is more than negative likelihood ratio, this means that the applied

fuzzy rules well performed when classified human skin images. Meanwhile, a value

of 0.0001 of P-value is obtained from the experiment which proved that the fuzzy

rules have the ability to classify between human and animal skin within images with

the value is significant low and different from 0.5.









4.4.2 Analysis of Fuzzy Logic Ability to Discriminate Porn and Non-Porn

Images







The classification result of porn and non-porn images is shown in Table C.2

which is listed in Appendix C shows. In Figure 4.8, the classification result of porn

and non-porn images is shown in ROC curve. The value of Area under ROC for

classification result on porn and non-porn images using fuzzy logic is only 5.5 which

is very unsatisfied and can be considered fail. The highest accuracy rate was

achieved when the criterion was 97 which the true positive rate was 68% and the true

negative rate is 98%. Although the positive likelihood ratio is only slightly better

than the negative likelihood ratio, it proved that fuzzy rules are capable in

discriminating porn and non-porn images. As stated in 3.3, a classification method

with p-value under the true Area under ROC curve (0.5) is considered good.

Therefore, the p-value of 0.2121 from Figure 4.8 which is still far below 0.5 also

provided evidence that this method proposed are able to discriminate between porn

and non-porn images.

58







100







Sensitivity: 85.0

80 Specificity: 39.0

Criterion : =0 100.00 0.00

>0 96.00 1.00

>0.6797 94.00 1.00

>1.818 94.00 7.00

>2.138 93.00 7.00

>5.3418 93.00 14.00

>5.418 92.00 14.00

>6.051 92.00 17.00

>6.2402 91.00 17.00

>6.3106 91.00 18.00

>9.134 89.00 18.00

>10.28 89.00 20.00

>11.58 86.00 20.00

>12.03 86.00 21.00

>12.05 85.00 21.00

>16.29 85.00 33.00

>17.1 84.00 33.00

>42.35 84.00 64.00

>45.51 81.00 64.00

>49.65 81.00 68.00

>49.89 80.00 68.00

>53.27 80.00 70.00

>54.32 78.00 70.00

>54.89 78.00 71.00

>56 77.00 71.00

>64.44 77.00 81.00

>65.92 75.00 81.00

>67.21 75.00 82.00

>68.55 74.00 82.00

>71.53 74.00 84.00

>71.6 73.00 84.00

>71.81 73.00 85.00

>72.13 72.00 85.00

>74.03 72.00 86.00

>74.05 71.00 86.00

>76.04 71.00 90.00

>77.38 70.00 90.00

>77.69 70.00 91.00

>80.14 68.00 91.00

>80.85 * 68.00 93.00

>89.31 63.00 93.00

>91.22 63.00 94.00

>93.78 62.00 94.00

>94.44 62.00 95.00

>95.02 61.00 95.00

>96.68 61.00 97.00

>99.94 46.00 97.00

>100 0.00 100.00

75





Table C.2: Classification Result of Porn and Non-Porn Images

Criterion True Positive Rate (TPR) False Positive Rate (FPR)

< 4.332 0.00 100.00

<=4.332 1.00 100.00

<=21.55 1.00 95.00

<=23.7 3.00 95.00

<=26.09 3.00 93.00

<=26.59 5.00 93.00

<=32.67 5.00 91.00

<=33.95 7.00 91.00

<=36.88 7.00 88.00

<=37.49 8.00 88.00

<=39.59 8.00 84.00

<=42.41 11.00 84.00

<=47.09 11.00 81.00

<=47.61 12.00 81.00

<=48.38 12.00 79.00

<=50.42 15.00 79.00

<=53.93 15.00 73.00

<=54.59 18.00 73.00

<=55.79 18.00 72.00

<=57.83 20.00 72.00

<=58.18 20.00 71.00

<=59.94 24.00 71.00

<=61.19 24.00 69.00

<=61.29 25.00 69.00

<=61.36 25.00 68.00

<=61.48 26.00 68.00

<=63.73 26.00 66.00

<=65.02 29.00 66.00

<=65.24 29.00 65.00

<=65.4 31.00 65.00

<=66.31 31.00 63.00

<=66.77 32.00 63.00

<=67.62 32.00 62.00

<=68.01 33.00 62.00

<=68.49 33.00 61.00

<=68.58 34.00 61.00

<=68.96 34.00 60.00

<=70.87 39.00 60.00

<=71.37 39.00 59.00

<=72.12 42.00 59.00

<=72.22 42.00 58.00

<=72.85 44.00 58.00

<=74.65 44.00 56.00

<=76 46.00 56.00

<=76.64 46.00 54.00

<=78.25 48.00 54.00

<=78.42 48.00 53.00

<=78.98 49.00 53.00

<=80.26 49.00 51.00

<=80.82 50.00 51.00

<=81.03 50.00 50.00

<=82.29 54.00 50.00

<=82.47 54.00 49.00

<=86.96 59.00 49.00

<=87 59.00 48.00

<=89.79 64.00 48.00

<=89.9 64.00 47.00

<=91.76 69.00 47.00

<=92.15 69.00 45.00

<=92.22 70.00 45.00

<=92.33 70.00 44.00

<=92.52 72.00 44.00

<=92.98 72.00 43.00

<=93.32 74.00 43.00

<=93.66 74.00 41.00

<=93.88 76.00 41.00

<=93.96 76.00 40.00

<=94.81 79.00 40.00

76





Criterion True Positive Rate (TPR) False Positive Rate (FPR)

<=96.61 79.00 39.00

<=97.64 * 85.00 39.00

<=97.65 85.00 38.00

<=97.66 86.00 38.00

<=97.87 86.00 36.00

<=98.24 87.00 36.00

<=98.32 87.00 34.00

<=98.45 88.00 34.00

<=98.49 88.00 33.00

<=98.73 90.00 33.00

<=98.75 90.00 32.00

<=98.88 91.00 31.00

<=99.23 91.00 28.00

<=99.74 95.00 28.00

<=99.79 95.00 27.00

<=99.83 97.00 27.00

<=99.93 97.00 20.00

<=100 100.00 0.00