ANEW ROBUST FACE DETECTION IN COLOR IMAGES Sanun Srisukand

Document Sample
ANEW ROBUST FACE DETECTION IN COLOR IMAGES Sanun Srisukand Powered By Docstoc
					                        A NEW ROBUST FACE DETECTION IN COLOR IMAGES

                                          Sanun Srisuk and Werasak Kurutach

                              Advanced Machine Intelligence Research Laboratory
                                    Department of Computer Engineering
                            Mahanakorn University of Technology, Bangkok, Thailand.
                                         sanun, werasak @mut.ac.th


                        Abstract                                  having both simple and complex background scenes. Skin
                                                                  color [2,3,6,9,11-14,18,20-22] has been widely used with
In this paper, we propose a novel approach for robust skin        shape analysis [6,13,18,21,22] and facial feature extraction
segmentation and face similarity measure. The proposed            [2,3,11,12,20], to speed up the face detection process and
skin segmentation is a method for integrating the chromi-         to reduce the false alarm rate. Although skin detection has
nance components of any color model. The goal of the              widely been investigated using so many approaches, i.e.,
method of skin detection is to select the appropriate color       statistical analysis [10,11], fuzzy pattern matching [12] and
model for verifying the skin pixel under different lighting       color region mapping [18,21,22], there are still many open
conditions and various types of skin color. The enhanced          problems to be explored. Neural networks (NN) [8,9,17,20]
Hausdorff Distance, called RAMHD, is used to measure the          have been proposed for detecting faces in grey images. The
similarity between the face edge and an elliptical model in       computational complexity is very high because neural net-
the skin area. This method is very robust to the occlusions       works have to process many small local windows in the im-
of the face edge. Finally, the results of face similarity mea-    ages. In addition, support vector machines (SVM) [7,16]
sure will be improved by updating the elliptical model. We        have also been applied to the face detection. The two
will show the performance of skin segmentation and face           methodologies use the same approach in training with the
similarity measure with real images.                              face and nonface database. The SVM performs slightly bet-
                                                                  ter and is approximately 30 times faster than the NN [7]. In
                   1. INTRODUCTION                                recent years, the Hausdorff Distance (HD) [1,6,21,22] has
                                                                  been proposed for detecting human face by measuring the
Automatically detecting human faces is becoming a very            similarity between the elliptical model and the face edge.
important task in various applications, i.e., human computer      The advantage of using the elliptical fitting technique is that
interaction (HCI) and video indexing. In HCI, computers           it is more robust against noise and changes in illumination.
can adjust their behaviors by knowing the user’s feeling              This paper presents a fast algorithm of face region detec-
through his or her facial expressions. In this application        tion based on color and shape analysis. Firstly, we will pro-
the location of face must known a priori, before the ex-          pose a novel method for integrating the well known color
pression of the face can be analysed. In video indexing,          models by using Bayesian estimation and fuzzy set based
the human faces must be discovered as fast as possible due        concept. Secondly, the shape analysis is performed by us-
to the large video database. Although several methods are         ing RAMHD, an enhancement of the conventional Haus-
currently used to perform the face detection, there are still     dorff Distance [4,5]. Finally, we will provide an efficient
many factors that make the face detection more difficult,          algorithm for updating the elliptical model. The algorithm
such as scales, locations, orientations (upright and rotation),   introduced here is faster and more reliable than others even
occlusions and poses. However, a face detection technique         in the case where the image contains multiple faces, failure
that can be used in any real time application needs to sat-       features and occlusions of the edge of faces.
isfy a number of important criteria: (1) All the true faces           The organization of this paper is as follows. In section
should be detected as fast as possible. (2) Face detector         2, we present our contribution in the area of face detection,
should be robust with repect to occlusion and failure of the      which includes our proposed robust skin detection based on
edge of face. (3) No false face should be detected. (4)           Bayesian estimation and fuzzy membership concept, a fast
All of faces should be found automatically without the re-        and accurate face similarity measure based on robust Haus-
quirement of system to adjust the parameters in the process.      dorff Distance and a formulation for updating the elliptical
(5) Face detector should discover the faces from images           model. Section 3 will show and discuss the experimental re-
sults of our face detection algorithm in real image. Finally,       Ô´Û      ½ ¾ µ is the a posteriori probability of Û , given
conclusions will be presented in section 4.                        chrominance components ½ and ¾ . To solve the problem
                                                                   of finding an appropriate color model, the maximum a pos-
                                                                   teriori (MAP) is used, and defined as
         2. FACE DETECTION ALGORITHM                                                                                            Å                               Ö Ñ Ü
                                                                                                                                                                      Û            Ô´Û                         ½          ¾µ


In this section, we will present our face detection algorithm                                                                                                                          Ô´           ½          ¾         Û µ ¡ Ô´Û µ
                                                                                                                                                                Ö Ñ Ü
in three parts: skin detection using Bayesian estimation and
                                                                                                                                                                      Û                                   Ô´         ½           ¾µ

fuzzy set concept, face similarity measure using RAMHD                                                                                                          Ö Ñ Ü
                                                                                                                                                                      Û Ô´                          ½          ¾         Û µ ¡ Ô´Û µ                                     (4)
and a method for updating the elliptical model.
                                                                      From eq. (4), the probability Ô´ ½ ¾ µ can be dropped
                                                                   because it is a constant independent of Û . The result of
2.1. Skin detection under different races and varying il-          MAP is Û that maximizes the probability of Ô´Û       ½   ¾ µ.
lumination conditions                                              This implies that we should select the chrominance com-
                                                                   ponents if the result from eq. (4) is Û ½ (skin). Thus, the
Color is a useful piece of information in computer vision          decision function can be defined as the followings:
especially for skin detection. The skin detection is an early
                                                                    Ô´                                  ½             ¾         Û½ µ ¡ Ô´Û½ µ                    Ô´         ½          ¾             Û¾ µ ¡ Ô´Û¾ µ µ                          ½          ¾   ¾ Û½        (5)
process used to avoid the exhausive search for faces. How-
ever, there are many color models that can be used for dis-         Ô´                                  ½             ¾         Û¾ µ ¡ Ô´Û¾ µ                    Ô´         ½          ¾             Û½ µ ¡ Ô´Û½ µ µ                          ½          ¾   ¾ Û¾        (6)
criminating color information in modeling skin color. In
[15], Liu has contended that there does not exist a single          The chrominance components in a given color model will
color space that is appropriate for all kinds of images. For       be selected only if ½ ¾ ¾ Û½ . Let us denote «
example, Hue (H) is not reliable for the discrimination task        «½ «¾ «Ò « ¾ « be the selected chrominance com-
when the saturation is low. Also, the distribution of skin
area is consistent across different races in the Cb and Cr         ponents, e.g. «       À Ë Ö    Á É . After the chrominance

color spaces. Therefore, when we use different color mod-          components of a given color model has been selected by us-
els under uncontrolled conditions, the skin detection will         ing Bayes theorem, the next process is to integrate them.
consequently result in different ways. Vandenbroucke et al.        If the selected chrominance components, «, are assumed
[19] propose an hybridge color space using “knock out” al-
gorithm. The best set of the three ones has been selected          to be independent, these chromaticity values can be sepa-
from 14 available color features for classifying soccer play-      rately verified. Let us suppose that Ñ Ò and Ñ Ü (see
ers. However, the accuracy of skin detection depends on            Figure 1 (b)) are the respective boundaries of the range of
both the color model and the method of skin classification.         human skin (defined by user). It can be observed that most
Hence, the challenge problem is how to select color models
that suit skin pixel classifications under different conditions,    of component values in any color space of a skin area are
i.e., different races and varying illuminations. Here, we pro-     distributed near the center of the range (i.e. ´ Ñ Ò · Ñ Ü µ ).
                                                                                                                        ¾
pose a novel approach for selecting the appropriate chro-          This behavior implies that these component values are im-
maticity values from a given color model based on prob-            portant information. Thus, we need a function that result in
abilistic reasonings. It is well known that the robustness
against changes in illumination can be accomplished if a           the highest value when the component values fall near the
color space efficiently separates the chrominance from the          center of the human skin range.
luminance information. So, in our system, the color model
(R, G, B) of images is transformed to the well known color                                               1
                                                                                                                                                                                        pαi (X=x)
representation systems. Then, the chrominance components
                                                                         Grade of Membership Function




of the transformed color models are used to classify the skin
pixel, by discarding the luminance value. The probability of
skin or nonskin given ½ and ¾ is then computed by using                                                 0.5



the Bayesian estimation:
                             Ô´                Û µ ¡ Ô´Û µ
            Ô´Û     ½   ¾µ
                                  ½    ¾
                                                             (1)
                                      Ô´   ½     ¾µ                                                      0
                                                                                                              0   100     200    300   400    500   600   700   800   900   1000
                                                                                                                                                                                   x        0
                                                                                                                                                                                                0         50       100     150        200    250       300   350
                                                                                                                                                                                                                                                                     X
                                                                                                                                                                                                                                                                   400



  ½  and ¾ can be defined as the chrominance compo-                                                                        a            b            c            d                                                            Xmin Xmax

nents of a given color model, i.e.             ¾ Ö      Ü Ý
                                                                                                                                             (a)                                                                                      (b)
 À Ë           Ö    £ £ Ù£ Ú£ Á É ½ ¾ Ë . Ô´ ½ ¾
Û µ represents the probability of ½ and ¾ given class              Figure 1: (a) Fuzzy membership function. (b) The distribu-
Û . Û ¾ Û½ Û¾ denotes the skin (Û ½ ) and nonskin (Û ¾ )           tion of skin area in 1-D histogram.
classes. Let Û½ ´ ½ ¾ µ and Û¾ ´ ½ ¾ µ be the 2-D his-
togram of skin and nonskin areas in the chrominance com-
ponents ½ and ¾ . ÆÛ is the total number of attribute val-               The membership function is employed in our approach.
ues in the 2-D histogram Û ´ ½ ¾ µ. The probability of ½                                                                                                         ¼                                                        Ü                 ÓÖ     Ü
and ¾ , given Û , is then defined as                                                                                                                               ¾´   Ü       µ¾
                                                                                                                                                                                                                                      Ü            ´ · µ
                                                                                                                                                                   ´          µ¾                                                                     ¾
                 Ô´ ½ ¾ Û µ         Û ´ ½ ¾µ              (2)                                                                                                    ½
                                                                                                                                                                          
                                                                                                                                                                        Ü   ¾´     µ¾                                     ´ · µ
                                                                                                                                                                                                                                                   Ü
                                      ÆÛ                                                                                                                                          µ¾
                                                                                                                  Ü
                                                                                                                  ´                            µ
                                                                                                                                                                 ½
                                                                                                                                                                             ´                                              ¾
                                                                                                                                                                                                                                  Ü                                      (7)
 Ô´Û µ is the a priori probability of class Û defined by                                                                                                                     
                                  Æ                                                                                                                              ½      ¾´Ü                          µ¾
                                                                                                                                                                                                    µ¾
                                                                                                                                                                                                                                  Ü                ´ · µ

                        Ô´Û µ È Û
                                                                                                                                                                         ´                                                                           ¾
                                                          (3)                                                                                                     ¾´  Ü          µ¾                                       ´ · µ
                                                                                                                                                                                                                                                   Ü
                                    ÆÛ                                                                                                                             ´            µ¾                                          ¾
                                                                                                               Direction of D1
    The parameters        and , where                   , are                                                                           Direction of D2            r2x2+ r 2y 2- r2ry2= 0
                                                                                                                                                                   y     x       x
adjustable to reflect the meaning of the designed fuzzy                                                    ry
membership function. The parameters and are defined                                                   r8          r1                                                     d'
                                                                                                                                                                         i
as the respective lower and upper boundaries of any skin                                                                                                  pk
                                                                                                                                                                                       ej
area in any color space. When Ü falls in the range from                                         r7                    r2                                      di

to , the function ´Üµ will result in the value 1. Ü becomes
                                                                                                                               rx
                                                                                                r6                    r3

less important information when it falls in the range from                                           r5          r4
  to or from to . The results from eq. (7) of each
chrominance component are then integrated by

Ë×   Ò ´Ü ݵ
               È   « ¬ ¬           ­      ­ µ ¡ ´« µ                                           Figure 2: Our new searching strategy
     ½            ´           ·                               × Ò   (8)
     ¼    ÓØ   ÖÛ ×


where Ë× Ò      ¢       ¼ ½ is the skin color likeness func-
                                                                          boundary. We try to find    which is closest to Ô along
tion. The parameters ¬ ¾ «½Å Ò «¾Å Ò        «ÒÅ Ò and ­ ¾
  «½Å Ü «¾Å Ü «ÒÅ Ü are the minimum and maximum                           the directions of ½ and ¾ . In order to accomplish
values of the range of human skin in selected chrominance                 that task, the elliptical model can be automatically
                                                 ´­  ¬ µ
                                                         ¢½               defined in terms of Ö Ü         ´ Ê Ñ Ü   Ê Ñ Ò µ ¾ and
                                                                                                      ¿
components, respectively. The parameter            ½¼¼
                                                                          ÖÝ       ´ Ê Ñ Ü   Ê Ñ Ò µ ¾1 .
                                                                                 ¿
is defined as the length of the ranges from to and from                                                      Then, the algorithm of
to . The parameter ´« µ is the weighting coefficient asso-                 our new measure, called Robust Automatic Minimum
ciated with chrominance component « .
                                                                          HD-RAMHD, can be carried out step by step as follows.
          ´«   µ   ÈÒÔ« Ô
                        ´         ܵ
                                       ܵ
                                            Ò     ´   «   µ   ½     (9)
                      ½   «   ´                                           step 1.              Let              ½   ½   ¾ ½ and         ½    ¾    ¿
                                                                                           ¼½ ¼¾ ¼¿                        ¼
                                                                                                                 be the interior and        ¼ ¾
 Ô« ´      ܵ is the probability generated by the 1-D his-                  ¾
                                                                                                   ¾                                                  ¾

togram of chrominance component « . Note that this his-                   exterior directions of searching. Initialize  and ¼ to Ô
togram is normalized by dividing all entries with the great-              (      ¼ Ô ) and to 1 ( ½).
est value. The threshold × Ò is defined for making a deci-
                                                                                              and ¼ , perform the following test: If
sion whether a color pixel is skin or not.
                                                                          step 2. At each
                         ½
                             Ò                                                    or ¼     , goto step 3. Otherwise, increase by 1
                   × Ò          Ô« ´     ܵ              (10)
                              Ò    ½
                                                                          (    · ½) and repeat step 2.

    After completing the skin detection process, we deter-                step 3. Compute the minimum distance
mine each region Ê       ´Ê Ñ Ò Ê Ñ Ü Ê Ñ Ò Ê Ñ Ü µ of
the skin areas by directly evaluating their projections on the              Ê ÅÀ           Ô
                                                                                           ´    µ
x- and y-axis. Then, the color pixels in the region Ê are
                                                                                    Ñ Ò´   Ô                    ¾          Ô     ¼           ¾µ       ´            ¾    µ      ´   ¼   ¾    µ
converted to a binary image by the method of gray scale                                              Ô           ¼         ¾                          ´            ¾    µ      ´   ¼   ¾    µ
conversion and edge detection. In each region, we generate                                           Ô                     ¾                          ´            ¾    µ      ´   ¼   ¾    µ
the sets of the edge                               ´Ü Ý µ ½                                                                                           ´            ¾    µ      ´   ¼   ¾    µ
                             ½ ¾ ¿
           to be used for verifying whether the area Ê con-                                                                                                                                     (11)

tains an oval shape or not.                                               repeat steps 1, 2 and 3 for all Ô                                       ¾È .
                                                                          step 4. Calculate the similarity measure
2.2. Face similarity measure under partial occlusion
                                                                           Ê ÅÀ        ´   È              µ
In this section, we will propose a robust method for face
                                                                                ½                                              ½                                             Ò ´È µ
similarity measure under partial occlusions. This technique
                                                                                                ½              È                                           Ô
                                                                                                                                                                        ·
                                                                                                                                                                             Ò´È µ
                                                                                                                                                                                                (12)
is based on a modified Hausdorff Distance combined with                          ¾    ÜÔ    Ò È  ´        µ          Ô ¾È            ´       Ê ÅÀ          ´        µµ
our new searching strategy as described in Figure 2.                       where Ò ´È µ and Ò´È µ are the respective cardinalities of
    From Figure 2,      ¾ ½ and ¼ ¾ ¾ , ½                 Ò,              the two point sets Ô ¾ È Ê ÅÀ ´Ô µ                   and È .
are two points in the xy-plane that lie on the straight                     Ê ÅÀ ´Ô µ measures the distance of each point Ô ¾ È
line ½ and ¾, respectively.          ½ and    ¾ are called                to its nearest neighbor in     in the directions of ½ and
                                                                            ¾ . If some points in È near some points in     , both ex-
the interior and exterior directions of searching. Let                                            È
                                                                          pressions ½ ÜÔ Ò ´È µ Ô ¾È ´ Ê ÅÀ ´Ô µµ and Ò ´´È µµ
                                                                                             ½
     ¾     be a point of the binary image, and È                                                                                 ÒÈ
  Ô½ Ô¾ Ô¿ Ô È Ô          ´Ù Ú µ ½              È      be a                   1 Note that we will provide an efficient algorithm for automatically ad-

set of ellipses each of which is a set of points on its                   justing the size of the elliptical model in the next section.
in (12) will result in an approximated value of 1. The term                         ÖÝ . Let ÖÜ and ÖÝ are the x- and y-radius before updated.
                                                                                              ÓÐ     ÓÐ

´Ò ´È µ Ò´È µµ is designed to prevent the false detection, oc-                      The new radius can be evaluated by
curing when the point Ô is too discarded by the func-                                                             ¼                                     ½
tion. In our system, it is assumed that the occluded areas                                                                 ½
                                                                                               ÖÜ
                                                                                                Ò   Û
                                                                                                        ÖÜ
                                                                                                         ÓÐ
                                                                                                              ·                              ÐÐ Ô× ´Þ   µ   (16)
of the edge of face are less than 35%. Thus, the expression                                                           Ò´       ½µ   Þ¾
´Ò ´È µ Ò´È µµ     ¼     is acceptable values. To justify the                                                                            ½
similarity measure Ê ÅÀ           È ¢            ¼ ½ , it can be                    and                           ¼                                     ½
shown that                                              È
¯ Ê ÅÀ ´È µ                                                                                                                                         Þ¼ µ
                                                   ½                                                                       ½
                           ½, if and only if Ò ´È µ        Ô ¾È                                ÖÝ
                                                                                                Ò   Û
                                                                                                        ÖÝ
                                                                                                         ÓÐ
                                                                                                              ·                              ÐÐ Ô× ´        (17)
                            Ò ´È µ ½,                                                                                 Ò´
                                                                                                                      Þ¼ ¾ ¾   ¾µ
  ´ Ê ÅÀ ´Ô µµ       ¼ and Ò´È µ
¯ contains an oval shape, if and only if Ê ÅÀ ´È µ                                   where Ò´ ½ µ and Ò´ ¾ µ are the cardinalities of the points
           Ò ´È µ ¼ , where
 ÓÚ Ð and Ò´È µ                    ÓÚ Ð ¾ ¼ ½ is the threshold                      sets ½ and ¾ , respectively. If the second terms of (16) and
used to adjust the sensitivity of the face similarity measure.                      (17) are both positive, the elliptical model will be enlarged.
    To solve the problem of comparing È to the portions                             Otherwise, it will be reduced. The elliptical model can be
of the set      (e.g. as occurs when faces are partly oc-                           recreated using an updated radius Ö Ü Û and ÖÝ Û . Let ÆÜ and
                                                                                                                           Ò        Ò
cluded or some features of faces are failed), we define the
    Ê ¼ ¼ function as                                                               ÆÝ denote ÖÜ Û   ÖÜ and ÖÝ Û   ÖÝ , respectively. We can
                                                                                                 Ò     ÓÐ         Ò       ÓÐ

                                                                                    find the best matching from Ê ÅÀ ´È Ò Û ¨ Ø µ, where
                             Ü
                             ´ µ
                                           Ü       Ü                                È Ò Û is an updated elliptical model, and Ø is ´Ø Ü ØÝ µ such that
                                           ¼       Ü                         (13)
                                                                                      Ü   ÆÜ ØÜ        Ü · ÆÜ and Ý   ÆÝ ØÝ
                                                                                       Å                Å               Å                   Å
                                                                                                                                          Ý · ÆÝ .
 where is the threshold defined for discarding any large
value of Ü. In our experiments, the parameters and ÓÚ Ð
                                                                                                    3. EXPERIMENTAL RESULTS
are set to ¾ ¿ and ¼ , respectively. This means that the point
Ô such that Ê ÅÀ ´Ô µ              is indicated as the point in                     In our experimentation we have employed 475 real images
the occluded areas, we will discard that point as a noise. If                       collected from various sources, i.e. internet and XM2VTS
the condition ´ Ê ÅÀ ´Ô µµ             holds for all Ô ¾ È , the                    database [16]. Each of which contains multiple faces with
distance from È to must be within the threshold . Thus,                             various sizes and different lighting conditions, most of
the notion of resemblance encoded by this distance is that,                         faces are in a complex background scene. Parts of some
the distance from set È to should be less than or equal to                          faces are occluded and some features of faces are failed.
 . This is an important advantage of our new measure, be-                           The database is composed of Asian, European and African
cause it discards the large error rather than combining them                        faces. We implement the proposed method on Pentium III
with the small one. In section 3, we will show the perfor-                          450 MHz with 128 Mbytes of memory.
mance of our new measure compared to others.                                            To generate the statistics of skin and nonskin. The two
                                                                                    best sets that most likely to the skin and nonskin areas
2.3. Updating the elliptical model                                                  should be carefully selected from the image database. The
                                                                                    skin areas are easily collected from the part of human body,
The accuracy of face similarity measure also depends on the
size of the elliptical model. To achieve the most appropriate                       i.e., face, hand and arm. However, the nonskin areas are
size of the elliptical model, Ö Ü and ÖÝ need to be updated                         not easy to gather, this is because the nonskin areas can be
using the ellipse function                                                          represented by everything in the image database excluding
                    Ûµ
              ÐÐ Ô× ´                  Ü Ýµ
                                 ÐÐ Ô× ´           ÖÝ Ü¾ · ÖÜ Ý¾   ÖÜ ÖÝ
                                                    ¾       ¾       ¾ ¾
                                                                             (14)   the skin area. Therefore, to collect a certain nonskin and
                                                                                    to reduce the nonskin data set. The nonskin areas that are
which has the following properties:                                                 most likely to be mistaken to the skin area are collected for
                         ¼         ´Ü ݵ       × Ò×     Ø       ÐÐ Ô×   ÓÙÒ ÖÝ      building the statistical informations. This collection can be
  ÐÐ Ô× ´   Ü Ýµ         ¼         ´Ü ݵ       × ÓÒ Ø       ÐÐ Ô×    ÓÙÒ ÖÝ         greatly helping to make a decision in the case where a pixel
                         ¼         ´Ü ݵ       × ÓÙØ×       Ø     ÐÐ Ô×  ÓÙÒ ÖÝ
                                                                             (15)   overlap between skin and nonskin classes.
    Refer to section 2.2, the idea of the directions of search-                         Our approach has three stages. In the first stage, the re-
ing can also be used to update Ö Ü and ÖÝ . The ellipse                             gions that may contain a skin color were extracted by the
is divided into 8 regions , i.e. Ö , ½                (see Fig-                     method in section 2.1. Then, the RAMHD will be used in
ure 2). Let È Å be a set of elliptical models at the po-
sition Å , where Å is the position of the best match-                               the second stage to verify whether or not the skin areas con-
ing of face similarity measure from section 2.2. The near-                          tain an oval shape. In our system, the facial feature does
est neighbor of a point Ô ¾ È Å can be found using                                  not neccessary to verify. This is because, the result that pro-
steps 1, 2 and 3 of the RAMHD algorithm. At each Ô in                               duces from section 2.1 is only the skin area. Hence, if that
the regions Ö¾ Ö¿ Ö Ö and Ö½ Ö Ö Ö the nearest neigh-
bor of those points that does not rejected by the func-                             skin area contains an oval shape, it is assumed to be the
tion are retained in ½         Þ½ Þ¾ Þ¿ Þ ½ Þ ¾ ½ and                               boundary of the human face. This can enormously reduce
  ¾
         ¼ ¼ ¼
        Þ½ Þ¾ Þ¿ Þ ¼ ¾ Þ ¼ ¾ ¾ , respectively. The points                           the computational complexity. Finally, the elliptical model
in the sets ½ and ¾ are respectively used to adjust Ö Ü and                         will be updated by using the method in section 2.3.
    The results of the first stage will subsequently effect the
accuracy of the face similarity measure. So, it is seem im-
portant that this stage must produces the results robustly and
reliably. From Figure 3, we have shown the result of our
skin detection when we apply to the images under different
lighting conditions. Figure 3 (a), (b) and (c) show the effect                                   (a)

of dark, normal and bright illuminations. The results show
on the right of Figure 3, one can see that the changing of
illuminations does not significantly effect to the results of
our skin detection. Moreover, we do not need to adjust the
range of human skin to suit each image. This is one of the
key advantages of our proposal, because we do not require                                        (b)

one to prespecify the boundaries of the range of human skin
for any particular image. Figure 4 (a), (b) and (c) shows the
results of our skin detection under different races, the result
is much more accurate and efficient enough for using in the
face detection stage. The example result of the method for
updating an elliptical model can be shown in Figure 5.                                           (c)

    In Figure 6, we present some results of our face detec-
tion algorithm. We have tested the proposed method on a           Figure 3: The results of skin detection under varying illu-
large number of data set. The label in the bottom left cor-       minations.
ner (T/F/N/E/C) gives the numbers of total faces (T), found
faces (F), not-found faces (N), false detection (E) and the
computation time (C) of face detection. Figure 6 (a), (b) and
(c) show the results of sample images with multiple faces
of different sizes. No false alarm occurs in this samples.
Moreover, all faces can be automatically discovered with-
                                                                                                  (a)
out requirement a user-defined parameter in the process. We
present the results of our face detection with occluded faces
in Figure 6 (d). In this case, the face detector still found
the locations of these faces. Figure 6 (e) shows the result of
section 2.1. In Figure 6 (f), it shows that the result of our
face similarity measure is very robust even so many features
                                                                                                  (b)
of the edge of face are disappeared.


  Table 1. The comparison of our algorithm with others.
                      Found    Not-Found   False      Time
The proposed scheme   92.36%   7.64%       18.22%     846ms
                                                                                                  (c)
M2HD [1]              52.95%   47.05%      26.47%     2063s
MEHD [4]              66.3%    33.7%       28.4%      2230s
HD [5]                35.3%    64.7%       147.05%    1218300s    Figure 4: The results of skin detection under different races.
YES+HD [6]            82.36%   17.64%      411.76%    56.75s
HSV+AMHD [21]         88.24%   11.76%      23.52%     773ms
AMHD [22]             64.71%   35.29%      32.35%     1853s



   Table 1 compares the accuracy and the computation time
of our approach with others’ [1,4,5,6,21,22] based on the
numbers of found faces, not-found faces, false detection
and computation time. The successful rate of the proposed
method is 92.36%, and the false alarm rate is also not bad.
The accuracy of YES+HD and HSV+AMHD is close to the
successful rate of our scheme. However, the techniques in         Figure 5: The results of face detection before (left) and after
[1,4,5,6,21,22] are less efficient than our technique. This        (right) updating the elliptical model.
                                                                                          5. REFERENCES
                                                                  ½ B. Takacs and H. Wechsler, “Fast searching of digital face libraries
                                                                 using binary image metrics”, IEEE Int. Conf. on Patt. Recog., pp. 1235-
                                                                 1237, August 1998.
                                                                  ¾ C. Garcia and G. Tziritas, “Face detection using quantized skin color
                                                                 regions merging and wavelet packet analysis”, IEEE Trans. on Multime-
   6/6/0/0/1020 ms
                                 (a)                             dia, Vol. 1, No. 3, pp. 264-277, 1999.
                                                                  ¿ D. Chai and K. N. Ngan, “Face segmentation using skin-color map in
                                                                 videophone applications”, IEEE Trans. on Circuit System for Video Tech,
                                                                 Vol. 9, pp. 551-564, 1999.
                                                                     D. G. Sim, O. K. Kwon and R. H. Park, “Object matching algorithms
                                                                 using robust hausdorff distance measures”, IEEE Trans. on Image Pro-
                                                                 cessing, Vol. 8, pp. 425-429, 1999.
                                                                     D. P. Huttenlocher, G. A. Klanderman and W. J. Rucklidge, “Compar-
                                                                 ing image using the hausdorff distance”, IEEE Trans. on PAMI, Vol. 15,
                                                                 pp. 850-863, 1993.
    3/2/1/0/980 ms                     2/2/0/0/760 ms                E. Saber and A. M. Tekalp, “Face detection and facial feature extraction
                           (b)                          (c)
                                                                 using color, shape and symmetry-based cost functions”, IEEE Int. Conf.
                                                                 on PR., pp. 654-658, 1996.
                                                                     E. Osuna, R. Freund and F. Girosi, “Training support vector machines:
                                                                 an application to face detection”, IEEE Int. Conf. on Comput. Vis. Patt.
                                                                 Recog., pp. 130-136, June. 1997.
                                                                     H. A. Rowley, S. Baluja and T. Kanade, “Neural network-based face
                                                                 detection”, IEEE Trans. on PAMI, Vol. 20, pp. 23-38, 1998.
                                                                     H. Ishii, M. Fukumi and N. Akamatsu, “Face detection based on skin
   2/2/0/0/890 ms
                                                                 color information in visual scenes by neural networks”, Int. Conf. on Sys.,
                                           1/1/0/0/516 ms
                    (d)          (e)                      (f)    Man, and Cyb., pp. 350-355, 1999.
                                                                  ½¼ H. Schneiderman and T. Kanade, “A statistical method for 3-D object
                                                                 detection applied to faces and cars”, IEEE Int. Conf. on Com. Vis. and
Figure 6: The example results of our proposed method.            Pat. Rec., pp. 746-751, 2000.
                                                                  ½½ H. Wang and S.-F. Chang, “A highly efficient system for automatic
                                                                 face region detection in MPEG video”, IEEE Trans. on Circuit System for
is because they give the results of much higher false alarm      Video Tech, Vol. 7, No. 4, pp. 615-628, 1997.
                                                                  ½¾ H. Wu, Q. Chen and M. Yachida, “Face detection from color images
rate. Moreover, in order to use any face detection in real
                                                                 using a fuzzy pattern matching method”, IEEE Trans. on PAMI, Vol. 21,
time applications, the response time needs to be within a        No. 6, pp. 557-563, June 1999.
few seconds. The result in Table 1 shows that our proposed        ½¿ J.-C. Terrillon, M. David and S. Akamatsu, “Automatic detection of
technique does meet the response time requirement.               human faces in natural scene images by use of a skin color model and of
                                                                 invariant moments”, Int. Conf. on Aut. Face and Ges. Rec., pp. 112-117,
                                                                 14-16 April, 1998.
                          4. CONCLUSIONS                          ½ J.-C. Terrillon and M. N. Shirazi, “Comparative performance of dif-
                                                                 ferent skin chrominance models and chrominance spaces for the automatic
                                                                 detection of human faces in color images”, Int. Conf. on Aut. Face and
We have presented a novel approach for face detection in a       Ges. Rec., pp. 54-61, 28-30 March, 2000.
complex background scene. A major aim of our approach             ½ J. Liu and Y.-H. Yang, “Multiresolution color image segmentation”,
is to improve the speed, the robustness, the accuracy and        IEEE Trans. on PAMI, Vol. 16, No. 7, pp. 689-700, 1994.
                                                                  ½ K. Jonsson, J. Matas, J. Kittler and Y. P. Li, “Learning support vectors
the reliability of face detection system. New techniques of
                                                                 for face verification”, Int. Conf. on Aut. Face and Ges. Rec., pp. 208-213,
skin detection and face similarity measure (RAMHD) have          28-30 March, 2000.
been proposed. In addition, the experimental results pre-         ½ K. K. Sung and T. Poggio, “Example-based learning for view-based
sented in this paper have proved that our proposed system        human face detection”, IEEE Trans. on PAMI, Vol. 20, pp. 39-51, 1998.
                                                                  ½ K. Sobottka and I. Pitas, “A novel method for automatic face segmen-
is more efficient than others. Some capabilities of this ap-      tation, facial feature extraction and tracking”, J. Sig. Proc. Image Commu,
proach can be clarified as follows. 1) The skin segmentation      Vol. 12, pp. 263-281, 1998.
is robust against different lighting conditions and various       ½ N. Vandenbroucke, L. Macaire and J.-G. Postaire, “Color pixels clas-
types of skin color. 2) The face similarity measure is robust    sification in an Hybrid color space”, Int. Conf. on Image Processing, pp.
                                                                 176-180, 1998.
against the occlusion and failure of the edge of face. 3) Our     ¾¼ R. Feraud, O. J. Bernier, J. E. Viallet and M. Collobert, “A fast and
face detection is highly accurate and reliable and has low       accurate face detector based on neural networks”, IEEE Trans. on PAMI,
false alarm rate. 4) Our face detection algorithm is quite       Vol. 23, pp. 42-53, 2001.
simple and, hence, has low computation complexity.                ¾½ S. Srisuk and W. Kurutach, “Fast detection of scalable and multiple
                                                                 human faces” Int. Conf. on Intell. Tech., Bangkok, Thailand, pp. 351-357,
    The proposed method can be applied in many real-time         December 2000.
applications, such as it can be used in the initial process of    ¾¾ S. Srisuk and W. Kurutach, “A new hausdorff distance-based face de-
the snakes based face tracking system [18] and indexing of       tection”, IEEE Int. Conf. on Artificial Intell. Sci. Tech., Hobart, Australia,
face images in video database [11, 20].                          pp. 203-208, December 2000.