Professional Documents
Culture Documents
3445
that fallacious skin region detection is trounced because of the fact that the Intensity component can be
proficiently. easily accessed in the YCbCr space. From the
The proposed face detection algorithm performs converted image the Y and the Cb components are
efficiently. The algorithm detects the valid skin removed there by retaining only the Cr Component.
regions and projects the face region out of the detected Through experimentation it is evident that all the skin
skin regions [Fig 1]. candidates are rich in red colour [Fig 2].
The complexity involved in computation is relatively It can be seen from the experimental results that along
more proficient when compared to that of the prior with the skin region some red valued non skin regions
developed methodologies [55] because of the fact that have also been highlighted. The Skin colour model is
the luminance information is excluded from the applied on red rich candidates to extract only the valid
computation. The pixel and the region based valid skin skin candidates. This model is tested on a database
colour extraction technique are applied only on the which cumulates people from different parts of the
bespoken chrominance information.. world having varied skin colours. The results obtained
exhibit an accuracy of about 99%.
2.1 SKIN ILLUMINATION COMPENSATION
MODEL 2.2 BOUNDARY DETECTION
In Order to detect human faces in an image, a good The detected skin region output from ICM is
starting point may be to identify any face like object in converted into the binary form and the boundary of
the image. One possible way to achieve it is to find each of the detected skin regions is detected and
any region in the image that is skin coloured. In our marked [Fig 3].
work, skin tone is considered for detecting faces
because of its efficiency and invariance to head pose, 2.3 SIGNATURE PLOT
orientation Effective retrieval of similar shapes from image
and size[44], [45]. Moreover detection using skin tone database is a complex and time consuming problem.
is much faster than processing other facial features In most of the cases referenced below it is a
[51]. Despite the fact that human faces have variation compromise between accuracy and expected response
in their skin colours [44], it is noted that such variation time. Direct comparisons of the images is insufficient
lies more in the intensity rather than the colour itself, not only because of the High Raw Data
that is to say that the skin colour chromaticity varies Dimensionality but also because of the shape
slightly between varied cases compared to the transformations (which occur due to different
intensity variation. This may have something to do viewpoints) can deteriorate the matching process and
with the fact that there are differences in the amount of thereby in turn may influence the face detection course
melanin in the skin, which affects the amount of light of action. In order to deploy and guarantee the high
absorption, but little or no effects on the spectral quality matching we have proposed a technique which
characteristics between individuals[40], [44]. deploys the Shape descriptors / Signatures of the
A Skin Illumination Compensation model can be boundary under consideration. Owing to our extensive
developed by combining the skin colour regions from study and ardent observation of the experimental
a set of face images with different skin colours. The results it can be stated that the Signatures adequately
appearance of the skin-tone depends on the lighting capture the shape characteristics and are typically
conditions. We introduce a illumination compensation tailored to be immune to transformations such as
model, where the captured image in the RGB colour translation.
space is converted into the YCbCr colour space The Signature of a detected objects’ boundary is
3446
plotted. The process is repeated until the signatures of faces (See Results), faces of any size, faces with
all the detected objects have been plotted. An glasses, intensity variations etc. Our algorithm can
exhaustive study of the Signatures of the various also detect the nonfrontal faces irrespective of the fact
object boundaries throw light on the fact that the that whether the eyes and nose are visible or not [42].
signatures pertaining to the face candidates will be We would like to claim the superiority of our
lesser than the signatures corresponding to the other approach over the approaches [6], [8], [9], [10], and
detected skin candidates [Fig 4]. [42] in Face Detection.
A summary of the detection results, are presented in
2.4 FACE CANDIDATE EXTRACTION the Table2. Subjects considered in our database belong
The valid face candidates are extracted by analyzing to several racial groups and lighting conditions. A
the signatures. Thus giving rise to the valid face detected face is a correct detection with a small
candidates [Fig 5]. amount of about 1% tolerance. The detection ratio is
computed by the ratio of the number of correct
3. EXPERIMENTAL RESULTS detections in a gallery to that of all the human faces in
Our algorithm is evaluated on several face image the gallery. The detection rate is around 99%, if the
databases, including photo collections (indoor and images considered have faces camouflaged with the
outdoor) and News Paper Photo collections. Face background then the face detection efficiency comes
databases readily available on World Wide Web, down to about 92%. The reason for this decrease in
designed solely for face detection and face the detection rate is due to the reason that the face is
recognition, including FERET Face database [15], merged with that of the background. The entire
[19], [34], [42] usually contain Gray scale images and database includes 225 images. One of the challenging
therefore, in our opinion are not suitable for evaluating episodes in face detection is detecting faces in groups
face detection algorithms. Most commonly, databases and our algorithm gives very excellent detection
used for face detection and face recognition include accuracy [Fig 6].
the images taken from Carnegie Mellon University
(CMU) database, MIT, FERET etc (See Table 1) 4. CONCLUSIONS
contain Gray scale images only. Because of this we We through this research paper are putting forth a
have made a humble effort to construct our own humble effort in the form of an algorithm to detect
database for face detection by collecting colour faces under varied conditions. Our detection algorithm
photographs in high degree of variability captured by initially takes the colour image and applies the skin
using CCD Cameras, Digital Cameras, the World illumination compensation model to modify the RGB
Wide Web, photographs from news paper websites to YCbCr converted image suitably to detect the valid
and personal photo collections. These colour images skin regions. The detected skin regions are converted
collected in our database are under varying lighting into the binary format and the signatures
conditions and with complex backgrounds. Further, corresponding to each of the detected skin regions is
the images considered in our database contain found and the skin region corresponding to the face
multiple faces with variations in colour, position, candidate is extracted by studying the signatures.
scale, orientation, 3D Pose and facial expression. With these results, we can confidently state that our
Our algorithm can detect multiple faces of different method outperforms all the published, popular
sizes with a wide range of facial variations in an methodologies.
image. Our algorithm works efficiently for occluded
3447
5. FIGURES AND TABLES stages involved in the Face Detection Process.
Figure 2 to Figure 6 in this section depicts the various
Step 3
Fig 2: Modified YCbCr Image with the ‘Y’ and ‘Cb’ components removed from Step 2 of Figure 1.
Fig 3: Illustration to show the Generation of all possible Skin Regions from Step 2 of Figure 1.
Fig 4: Signatures of all possible Skin Regions generated in the Step 3 of Figure 1.
Fig 5: Valid Face Candidate Extracted from the Detected Skin Regions from Step 4 of Figure 1.
Fig 6: Experimental results for different cases Considering all Extraneous Conditions like Illumination Variation,
Complex Backgrounds, Skin Tone Variation, and Occlusion with different poses for Multiple Faces
3448
Sl No Image Type (Both Indoor and No of Images in the Total No of Correct Face
Outdoor) Gallery Detections
01. Single Face with Complex Background 30 30
02. Multiple Faces With plain Background 40 40
03. Multiple Faces With Complex 65 65
Background
04. Single Face with very poor illumination 25 25
05. Multiple Faces with poor illumination 20 19
06. Single Occluded Face 10 10
07. Multiple faces with few faces occluded 31 30
TOTAL IMAGES = 225 TOTAL CORRECT DETECTIONS
= 223