INTRODUCTION
Face processing systems are not mere mechanisms to process
faces but rather mechanisms for deriving fine-grained discriminations
between visually similar exemplars. Face detection and tracking are the
foremost steps in face perception, since the knowledge about spatial features
of a face can be used very effectively for providing a wide variety of
information such as personality, social class level, age, health, gender,
etc. Most of the feature extraction algorithms purposeful on a universal
account will incline to produce false targets, as they are ineffectual
to resolve the face locality and for this reason they are impracticable.
Faces represent complex, multidimensional, meaningful
visual stimuli. Faces differ from other objects in several vital facets.
By and large it is enough to identify common objects at the basic level
(e.g., Chair, Car, Pen), without determining the specific exemplars of
the category (my chair, joe`s car, edi`s pen), but for faces it is necessary
to proceed beyond the general category face for determining the identity
of the said individual and non-homogeneity further compounds the problem.
Besides faces share the same basic configuration unlike object recognition.
Relatively high discriminability between numerous faces is achieved only
by small displacement of face features so that extremely reliable discrimination
occurs within the class of face objects. Substantially, a domain general
account for face perception is not possible.
In addition, behavioral studies also suggest that domain
specific mechanism helps to reduce the complexity of the surface whose
properties are decisively dependent for the accuracy of curvature estimation
to accomplish a precise throughput. Nevertheless, biological studies show
that regions in the fusiform gyrus is activated not only when subjects
view faces, but also atleast twice as strongly for faces as for a wide
variety of non face stimuli, including letters (Mario, 2005).
From these grounds, mechanisms to recognize individuals
are face-specific mechanisms making fine-grained discriminations between
exemplars and their accuracy at discriminating is higher, only when the
entire face is presented than including the other subjects. Other findings
put forward that face recognition involves special mechanisms that are
more inversion-sensitive and more holistic (less part based) than those
involved in object recognition.
The human face is a dynamic object making face detection
a intricate apprehension in computer vision, for which ample variety of
techniques have been recommended in the literature ranging from simple
edge-based algorithm to composite high-level approaches utilizing advanced
pattern recognition methods.
Current techniques employed in the task of face detection
are classified into two distinct groups` namely feature-based approaches
and image-based approaches. Feature based approaches relies on the application
of face knowledge and employs known properties of facial features. Representative
techniques of feature-based approaches include color-analysis (Cahi and
Nga, 1999; Michael and James, 1998), feature searching (Leung et al.,
1995; Liu et al., 2002) and template matching (Craw et al.,
1992; Lanitis et al., 1995). In contrast, the image-based approaches
exploit pattern recognition theories and make use of training algorithms
to implicitly incorporate face knowledge into the system. Typical techniques
using this approach include linear subspace method, such as Eigen faces
(Turk and Pentland, 1991), neural networks (Agui et al., 1992;
Henry et al., 1998) and various statistical methods (Erik and Boon,
2001; Colmenarez and Huang, 1997). The intended face detection algorithm
presented in this study uses a permutation of feature-based approaches
to maximize the probability of detection.
SKIN REGION SEGMENTATION
The inspiration to use skin color analysis for initial
classification of an image into probable face and non-face region stems
from a number of simple but powerful characteristics of skin color.
| • |
Color of human skin is different from the color of most other natural
objects in the world. |
| • |
Under certain conditions, skin color is orientation invariant. |
| • |
It is less obscure to process skin in contrast to other facial features. |
Skin color detection can be performed via computationally intensive iterative
methods such as skin patch merging (Garcia and Tziritas, 1999), dynamic link
architecture (Laddes et al., 1993), elastic graph matching (Wiskott et
al., 1999) etc. However, real time applications heavily rely on the fast
pixel based approaches rather than the above optimization methods to meet stringent
time requirements. The proposed pixel based approach for skin tone detection
is a modified and extended form of Fleck and Forsy algorithm (Mario, 2005) that
performs reasonably well both in terms of skin color detection and speed.
Color space: One imperative factor that should
be decided in skin color detection is the color space to operate on, in
order to effectively segment color of different shades under a variety
of lighting conditions, since color information can significantly simplify
the task of face localization in images with complex background. The images
are conventionally represented in the RGB domain for display, where the
color and intensity components are correlated. This requires some preprocessing
in order to reduce the effects of unwanted variations in intensity (Michael
and James, 1998), The color information expressed in the RGB color space
has not provided any improvement in the detection rate. Using R, G or
B independently provides the same results as using R, G and B simultaneously
and exactly the same results as using the luminance (Y) only. Also the
RGB components are subjected to the lighting conditions and thus face
detection may fail if there are frequent changes in the lighting conditions.
Besides, a color space with a separation of the luminance
and chrominance information`s tend to provide better results than a color
space with these information`s mixed (as RGB). Consequently, the original
RGB image, which incorporates luminance information, must be transformed
into a color space that separates the intensity and color information
of the image respectively. Various studies have shown that the skin color
regions are highly correlated in their hue and saturation channels or
chrominance channels (Cahi and Nga, 1999). Hence the input color image
typically in the RGB format is usually converted into color components
such as HSV, YCbCr, YIQ formats in the color space.
The information of different shades is usually present only in the intensity
channel (Y or V) and it captures variations in lighting conditions. Therefore
CbCr or HS channels are used in skin color segmentation.
Skin segmentation: In this approach, the RGB image
is transformed to log-opponent (IRgBy) values from
which the texture amplitude, hue and saturation are computed (Jay, 1997).
The conversion from RGB to log opponents is calculated as:
| I |
= |
[L(R) + L(B) + L(G)]/3 |
| R |
= |
L(R)-L(G) |
| B |
= |
L(B)-[L(G) + L(R)]/2 |
The L(x) operator is defined as L(X) = 105* log10
(X+1). A texture amplitude map is used to find regions of low texture
information. Skin in images tends to have very smooth texture and so one
of the constraints on detecting skin regions is to select only those regions
with little texture. Hue and saturation are approximated to select those
regions whose color matches that of skin and the conversions from log
opponent are given as:
| Hue |
= atan2 (Rg, By) |
| Saturation |
= sqrt (Rg2 + By2) |
Employing constraints on texture amplitude, hue and saturation
marks skin regions.
 |
| Fig. 1: |
(a) Input image (b) skin marked image |
|
| Fig. 2: |
(a) Skin connected component, (b) face skin connected
component using compactness, (d) face skin connected component using
orientation, (e) face skin connected component using solidity and
(f) resultant face component |
If a pixel falls in a particular range it is marked as being skin in a binary skin array,
where 1 corresponds to the coordinates being a skin pixel in the original
image and 0 corresponds to a non-skin pixel. The binary skin map region
is then processed using morphological operations to enlarge the skin map
region to include skin and background border pixels (i.e.,) regions near
hair or other features (Fig. 1).
Face skin connected component operator: The skin
filter described earlier tends to produce false skin regions, since there
is a tendency for highly saturated reds and yellows to be detected as
skin and this in turn calls for the use of face skin connected component
operators to eliminate those false regions (Fig. 2).
Connected component operators are non-linear filters that eliminate parts
of the image, while preserving the contour of the remaining parts. This
simplification property makes them attractive for segmentation and pattern
recognition applications.
Connected operators use certain decision criteria`s to
either retain or eradicate a connected component without affecting the
other components. Three shape-based connected operators, Compactness,
Orientation and Solidity are applied over the skin components to decide
whether they represent a face or not. The simple but effective decision
criteria or operators make use of basic assumptions about the shape of
the face to remove unwanted non-face components and are computed using
the perimeter P and the size Dx and Dy of the min-max
box of the connected components as configured below.
| • |
Compactness of a connected component is defined as the perimeter
of the region. |
The criterion is maximized in case of face objects for
which a suitable threshold is fixed based on the observations of various
face components. Significantly, components exhibiting high value for this
operator are selected as face component.
| • |
Orientation is but the aspect ratio of the min-max box surrounding
the component |
Based on the observation of a number of images it is
assumed that normal face components have orientation within a specific
range. Using this criterion the skin region containing the face is cropped
accordingly reducing the size of the candidate face region.
| • |
Solidity of a connected component is defined as the ratio of its
perimeter to the perimeter of the min-max box. |
It gives a measure of the perimeter occupied by the connected
components with respect to its min-max box dimensions. The solidity operator
also assumes a high value for face components. If a solidity of a component
is lesser than a specified threshold value, it is eliminated otherwise
retained.
Thereby the first phase deploying a skin filter together
with the face skin connected component operators comprehensively isolates
the potential face region from a complex background.
ENTRANT FACE SEGMENTATION
The skin filter may prove to be a generalization of the
problem of face detection. To make the algorithm more robust, some additional
constraints on the measurement of the face with respect to its feature
can be imposed to segment the probable face region required for face perception.
Detection of the entrant face region is itself a 2-stage process which
primarily extorts a feature point from the skin region, subsequently followed
by exercising anthropometrical statistics of a geometric model to automatically
isolate the implied face.
Eye map: Detecting the eyemap is based on the
spatial arrangement of the Chrominance and Luminance components in the
YCbCr color space (Reim-Lien et al., 2002).
Among the various facial features, eye and mouth are the most prominent
features for segmentation in color space although the mouth map is indistinct
especially in dark and light tone faces, since the red component is weaker
in such cases. The Chrominance components contained in Cb and
Cr and the luminance component contained in Y is employed to
build two separate eyemaps (Fig. 3). These eye maps
are then merged into a single map by an AND operation.
The resulting eyemap is then dilated, masked and normalized
to brighten the eye region and suppress the other facial areas. Erosion
removes small and thin isolated noise, while dilation preserves the required
components. Normally the structuring element operational should not be
more than that of the smallest eye which possibly will leave more connected
components other than the eye. This again calls for the connected component
operators to extort the contour of the eye from the isolated components
while retaining the shape of the eye components (Fig. 4
).
Together with the compactness and solidity operators
that assume low values for eye component, a symmetric operator is also
practical to retain the symmetric component of the eye. The components
that fall out of symmetry are eliminated, thereby reducing the number
of components and it is these retained symmetric components that enfold
the eye.
Anthropometrics: In human-factor analysis, a known
range for human measurements can assist in deriving the probabilistic
face region for face perception. Anthropometry is the biological science
of human body measurements in which the rich description of human geometry
developed provides qualitative information about the shape and proposition of faces (Farkas, 1994;
DeCarlo et al., 1998).
 |
| Fig. 3: |
(a) EyeMapC AND EyeMapL, (b) EyeMap and (c) Eye component |
 |
| Fig. 4: |
(a) and (b) eye map components including noise and
(c) segmented eye component |
 |
| Fig. 5: |
Probable face region |
The theory states that all facial features
are proportional to each other and the proportions of the face are as
follows.
Hface = 1.8 DEye; HEye
= 1/3 HFace; WEye = 0.2225 HFace |
The height of the face defined as a measure above the
eyebrows to just below the face, can be determined from the dimension
of the eye furnished during the construction of the eye components. Using
the above aspect ratios, the rationale behind localizing and extracting
the characteristic face region from the derived skin region is accomplished
(Fig. 5 ).
RESULTS AND DISCUSSION
The efficiency of the new-fangled face detection system
discussed was analyzed using a photo gallery of 25 frontal color images,
which was collected from known individuals taken from different cameras under different
lighting conditions (Fig. 6).
 |
| Fig. 6: |
Extracts from the photo gallery |
The selection of the color
space and the skin segmentation operators has proven to identify the face
component precisely except for an overestimation of the potential skin
objects which habitually become apparent when people appear partially
clothed. The resultant being a very large skin map containing the entire
head arms or torso is abridged by the three skin connected component operators
designed to crop the face component. This reduces a lot of computational
overhead for the second phase of segmenting the characteristic face. The
chrominance and luminance information easily de-embedded from the YCbCr
color space together with the component operators is utilized to
retain the shape of the eye components while removing the other connected
components. The random measurements from the eye region employed to smoothly
populate the plausible face according to anthropometrical statistics extract
the face region containing only the facial feature necessary for face
perception.
The proposed method extracts features relative to the
location of the eyes and hence it is invariant to size and tilt of the
face. But in case of occlusions like spectacles the quality of the spectacles
can also contribute to an engorged face region, more than desired as in the last
example in Fig. 6, nevertheless it adds to success.
The human skin color modeling using IRgBy color mode however has serious
problems when the algorithm loses track of the face, which occurs when
the face does not occupy a significant area of the image resulting in
a false negative. The intended system is efficient in terms of computational
power and memory requirements and experimental results prove that the
system gives an average deduction rate of 96% in real time under natural
lighting conditions (Table 1).
CONCLUSION
The major intent of this study, to provide a rigorous
cognitive process using holistic and configural perception has been achieved
from experimental manipulation of faces. Face detection algorithm using
holistic has the advantage of finding small faces, in addition to segmenting
faces in poor quality images. Detection using geometrical facial feature
provide a good solution for detecting faces in different pose. A combination
of holistic and feature based approaches have proven to be very promising
from the above discussions.
Defining additional features like mouth to appreciate
face dimension computations in anthropometrics can further enhance face
segmentation. Besides the system can be extended to deduct multiple faces
with variation in color, position, scale and orientation.
Subscribe Today
