International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print), ISSN 0976 – 6456(Online) Volume 6, Issue 1, JanuaryApril (2015), pp. 0107 © IAEME
1
A RESEARCH STUDY ON RECENT SKIN COLOR BASED STATISTICAL SEGMENTATION MODELING TECHNIQUES
Noor Adnan Ibraheem
1, 2
, RafiqulZaman Khan
2 1
Department of Computer Science, College of Science for Women, University of Baghdad, Baghdad, Iraq
2
Department of Computer Science, Faculty of Science, Aligarh Muslim University, Aligarh, India
ABSTRACT
The fundamental thought behind pixel based skin color segmentation is to detect and extract the pigment color of human’s skin from the input image since it may contain other like skin color. The skin color is a perfect sign to the presence of human skin in a specific scene and a helpful indicator to minimize the search area to include the hand and face only. Skin color modeling techniques have been widely utilized for this purpose, the selection of color modeling techniques depends on the nature of the application with the suitable color model. In this paper we will discuss some recent skin color based segmentation modeling techniques applied to located and extract the hand gesture in order to prepare the segmented hand for the later gesture recognition stages.
Keywords:
Segmentation techniques, color space, Gaussian mixture model, histogram, skin color segmentation, color based segmentation.
1.
INTRODUCTION
To recognize hand gestures, the input image has to be subjected into several processes firstly. The most effective step is to isolate the hand gesture from the background and other unwanted objects; this isolation is usually performed by applying some of the segmentation
INTERNATIONAL JOURNAL OF GRAPHICS AND MULTIMEDIA (IJGM)
ISSN 0976  6448 (Print)
ISSN 0976 6456 (Online) Volume 6, Issue 1, JanuaryApril 2015, pp. 0107
© IAEME: www.iaeme.com/IJGM.asp Journal Impact Factor (2015): 4.5273 (Calculated by GISI) www.jifactor.com
IJGM © I A E M E
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print), ISSN 0976 – 6456(Online) Volume 6, Issue 1, JanuaryApril (2015), pp. 0107 © IAEME
2
techniques. For the segmenting of hand gestures in the vision based method, there is two different issues should be determined; the segmentation of spatialgesture or the segmentation of temporalgesture [1]. The spatial gesture segmentation concerned with deciding the place where the gesture happens while the temporalgesturesegmentation related with the time when it happens. Usually previous researches used to perform one of these methods; however, recently attentions are focused towards applying both methods when dealing with the gesture recognition system [1][2]. When dealing with the hand gesture, the most effective parameter that refers to the existence of the hand; is the color of the skin, and the challenges here is how to segment the hand from other objects that are similar to human skin color. In this case the selection of proper color modeling intervenes with the help of suitable segmentation algorithm to solve this problem and extract the hand gesture correctly. In most modeling operations, the lighting or illumination component is neglected to ensure the perfect performance of the color model that is unaffected with sensitivity to illumination changes[3][4]. For gesture recognition systems, the segmentation process considered the most decisive step which demanded an accurate and robust consideration [5].The most effective cue that can be applied; is the pixel color where each pixel can be classified into skin and nonskin pixel depending in pixel’s intensity which provide the ability to recognize the pixel without regarding object size, orientation, and can reduce the search space. To construct any human skin color segmentation we have to consider to important issues; 1) the selected color model, and 2) modeling technique applied[5][6]. Various segmentation techniques have been utilized for pixel based skin color detection, these techniques include;generic skin model, Statistical color modeling techniques which can further subdivided into different algorithms that comprises; Explicitly defined skin region, single Gaussian Model (GM) and Gaussian Mixture Model (GMM), histogram based lookup table (LUT), and Bayes classifier [5][6]. In this paper, we will discuss two different kinds of statistical techniques that are recently used for skin color classification which are GMM and histogram based lookup table (LUT), with comparisons between these methods.
2.
SKIN COLOR MODELING TECHNIQUES
In order to discriminate the skin color pixel, pixel based color modeling techniques should be used to identify skin and nonskin pixels[7][8], these techniques have been widely applied for shape and color modeling and it comprise; explicitly defined skin region, and Statistical modeling techniques which in turn contains; parametric Color Distribution Modeling, and Nonparametric Color Distribution Modeling. In Statistical modeling techniques, the training data are used to model skin color distribution [7]:
A.
Parametric Color Distribution Modeling
This technique constructs the model by approximating the trained data (which can be small size) to classify the pixel into skin and nonskin pixel, the distribution shapeof the trained data is represented in complex form [3]. In this model, the parameters are initialized and estimated using statistical iteration algorithms such as kmeans clustering and Expectation Maximization algorithm [4]. The main benefit of the model is that it relies on the data distribution shape with no need for large storage space, but it suffers from long
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print), ISSN 0976 – 6456(Online) Volume 6, Issue 1, JanuaryApril (2015), pp. 0107 © IAEME
3
computational time required during training and running process, an example of this modeling technique is GM, and GMM.
B.
Nonparametric Color Distribution Modeling
In this model, the probability of color pixel is constructed from the training data with no require to fit the data in a modeling technique or for parameters initialization and estimation [4]. This technique characterized by speed in implementation, robust, and the distribution shape has no impact on it, but it suffers from the effect of illumination changes and it considered memory space consumption. Histogram with Bayesian classifier, Normalized lookup table (LUT), and Self Organizing Map (SOM) techniques are examples of this method [4].
3.
RELATED SKIN COLOR MODELING TECHNIQUES
With the development in hand detection techniques, it still represents a crucial step in machine vision field, the most effective challenges that might erases are; illumination variations, objects occlusions, low resolution and low quality video for tracking purposes[9][10]. Alon et al. [1] suggested an algorithm that take the most plausible hand gesturing from a set of candidate hand positions for every frame of the input gesturing sequence, this algorithm will minimize the time required for the preprocessing/ segmentation process and hence speed up the algorithm [1]. Alvar et al. [11] introduce a Mixture of Merged Gaussian Algorithm (MMGA) that merged Mixture of Gaussians probability model and the Real Time Dynamic Ellipsoidal Neural Network (RTDENN) learning model in order to minimize the required time for real time applications, the output results show the robustness of the suggested algorithm according to illumination variations. Chen et al. [12] detect the hand gesture by firstly applied the background subtraction method and then converted the output result into a binary image [12]. In [3], the authors applied multiple GMM (three GMM for the RGB, HSV, and YCbCr color models respectively) for hand segmentation to form (MuGMM) model using the following formula:
 = max
∈
;
,∑
(1) Where
represent the number of color models applied,
,∑
represent the papramters related to the color model m for each single GMM,
;
,∑
represent the popular GMM function, and

is the probability of the color pixel c being a skin color pixel. Whereas in [5], we applied the histogram modeling technique instead of GMM, the RGB histogram divides the RGB components into eight number of bins for each component (
8×8×8
) histogram of 512 bins, The HSV histogram divided into eighteen regions for (H) component and three regions for the S, and V components, this provide (
18×3×3
) histograms of 162 bins, and the YCbCr histogram divides the (Y) luminance component into eight regions, and (Cb, and Cr) components into four regions, which form
8×4×4
histogram of 128 bins; only the chrominance components have been used (RG, HS, and CbCr) to reduce the illumination effects by ignoring the luminance factor [5].
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print), ISSN 0976 – 6456(Online) Volume 6, Issue 1, JanuaryApril (2015), pp. 0107 © IAEME
4
For a single histogram for a color model, the probability distribution
has been defined by:
ℎ =
! "!
(2) Where
represent the histogram value of bin for the color vector
, and
is the sum of all histogram bin values. The proposed modeling system is
 = max
# ∈ $
ℎ
#
(3) Where
%
is the number of the three color models used,
ℎ
#
is the histogram probability of the
color model, and

is the probability of color pixel
being a skin color [5]. In [6], authors applied a mixture of GMM (MiGMM) for the same set of color models, and the classification rate CR (defined in equation (5)) has been used as the weight for each color model.
& ' = ∑ (
&;
,∑
)*
(4) Where
&;
,∑+
, c, M, m are the parameters defined in Gaussian classifier,
is the weight of m parameter and it considered as the effective factor that will define the classification rate of the suggested model [6].
=
.
/
∑ .
002
, + = 1,5,6,
(5) In [4], we applied the same weights as an evaluation to calculate the probability of mixture of histogram techniques (MxHT) The mathematical representation of the MxHT is shown in the following equation
& ' = ∑ (
#
ℎ
#
,
7#)*
= 1,5,ℎ
(6) Where Wi is the weight’s value of the ith is the color model which computed by dividing the CR of the ith selected color model by the CRs of all the used color model as explained in equation (5)[4]. a b c d
International Journal of Graphics and Multimedia (IJGM), ISSN 0976 – 6448(Print), ISSN 0976 – 6456(Online) Volume 6, Issue 1, JanuaryApril (2015), pp. 0107 © IAEME
5
e f g h
Figure 1:
a, b, c, and d represents the srcinal testing images. e: MiGMM[6], f: MuGMM[3], g: MHT [5], h: MxHT[4]. The evaluation metric used to evaluate the performance of the modeling techniques consists of three metrics; CDR, FDR, and CR. These metrics can be defined as follows [3][5][4]: Correct Detection Rate (CDR): is the number of pixels that are correctly classified as skin pixel by the algorithm.
9: =

<
>
<?
×1@@A
(7) False Detection Rate (FDR): refers the number of pixels that are wrongly classified as nonskin pixel by the algorithm.
B: =
C
D<
>
D<?
×1@@A
(8) Classification Rate (CR): indicates the number of skin pixels that are correctly classified by the algorithm and ground truth divided by the maximum value from either the number of skin pixels classified by the algorithm or the number of skin pixels classified by the ground truth.
9 =

<
EFG>
<
,>
<?
H
(9) Where
I
J
is the total number of pixels that are classified as skin pixels correctly by the algorithm.
K
L
represents the total number of pixels that are correctly classified as skin pixels by the ground truth.
M
!J
is the total number of pixels that are wrongly classified as nonskin pixels by the algorithm.
K
!L
is the total number of pixels classified as nonskin pixels by the ground truth, and
K
J
represent the total number of pixels classified as nonskin pixels by the algorithm.in table 1, the three mentioned metric have been calculated for the sosme of the discussed skin color modeling algorithm.