IRJET- Survey on Face Detection Methods

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 11 | Nov 2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 1467
SURVEY ON FACE DETECTION METHODS
Sharmeena Naido1, Rosalyn R. Porle2
1,2Faculty of Engineering, Universiti Malaysia Sabah, Jalan UMS, 88400 Kota Kinabalu, Sabah, Malaysia.
-------------------------------------------------------------------------***------------------------------------------------------------------------
Abstract - Face detection has attracted attention from
many researchers due to its wide range of applicationssuch as
video surveillance, face recognition, object tracking and
expression analysis. It consists of three stages which are
preprocessing, feature extraction and classification. Firstly,
preprocessing is the process of extractingregionsfrom images
or real-time web camera, which then acts asafaceornon-face
candidate images. Secondly, feature extraction involves
segmenting the desired features from preprocessed images.
Lastly, classification is a process of clustering extracted
features based on certain criteria. In this paper, 15 papers
published from year 2013 to 2018 are reviewed. In general,
there are seven face detection methods which are Skin Colour
Segmentation, Viola and Jones, Haar features, 3D-mean shift,
Cascaded Head and Shoulder detection (CHSD), and
Libfacedetection. The findings show that skin colour
segmentation is the most popular method used for feature
extraction with 88% to 98% detection rate. Unlike skin colour
segmentation method, Viola and Jones method mostly
comprise of face regions and other parts of human body with
80% to 90% detection rate. OpenCV, Python or MATLAB can
be used to develop real-life face detection system.
Key Words: Face Detection; Haar features; Skin Colour
Segmentation; Viola and Jones
1.INTRODUCTION
In this 21st century, advancement in computer technology
has facilitated interest in human-computerinteraction(HCI)
system. Face detection is a fundamental procedure for any
HCI system. It can be considered as one of the most
challenging areas of research yet applicable in various areas
such as security, gesture analysis and biometrics. Variations
in pose, lighting condition, occlusion, facial expression,
orientation, scale and location are some of the challenges
associated with face detection.
Face refers to an individual’sfront part of thehead,which
includes areas of the forehead to the chin. It is a vital part of
the human body as it represents information such as
expression and identity [1]. Detection here refers to the
process of identifying the existence of individuals.
Four major categories of face detection methods are
knowledge-based, appearance-based, feature-based and
template matching [2]. Knowledge-based method contains
set of rules to detect face depict by human knowledge, for
instance, facial features include nose at the centre, mouth
under the nose and a pair of symmetric eyes. Appearance-
based method uses a face model based on analysis perform
on a set of training samples. Feature-based method aims to
detect faces by extracting facial features. Although this
method is flexible and invariant to changes in pose and
orientation, it is easily affected by noise, illumination,
occlusion and other environmental factors. Template
matching method correlates input and template images for
face detection. This method is simple, but rotation, direction
and size are some of the factors that can affect the
effectiveness of one’s system.
Over the years, there are numerous researchdoneonface
detection to improvisean existing facedetectionsystemorto
propose a new algorithm for face detection. The primary
purpose of this paper is to provide insight on some of the
well-known methods used for the feature extraction and
classification process.
2. REVIEW OF FACE DETECTION METHODS
2.1 Skin Colour Segmentation
The skincoloursegmentationmethodinvolvesseparation
of skinand non-skin pixels. RGB, YCbCr and CbCrare someof
the colour space used for skin colour modelling.
Face detection using CbCr colour space in video is
presented in [3], which aimed to have a negligible
misclassification rate. This colour space was used for human
skin colour is less dependent to brightness. Experimentation
on the performance of face detectionwascarriedouton4000
images of Ytcelebrity, YouTube and FJU database. Images in
the video comprise of only face region. The detection
accuracy achievedwas98%and95%,whereasfalserejection
and acceptance rate obtained were approximately 10%.
Face detection using YCbCr colour space is presented in
[4]. The YCbCr colour space was used for skin colour
detection and segmentation because this colour space has
been encoded in most video media. Experimentation on the
performance of face detection was carried outon150images
which comprise of only face region. 95 imageswerecorrectly
detected, and 55 images were falsely detected due to low
image quality or face size below than 32×32.
In [5], face detection using a combination of motion and
skin colour segmentation is presented. YCbCr colour space
was used to perform skin colour segmentation.
Experimentation on the performance of face detection was
carried out on seven different ChokePoint video datasets,
each consists of 200 frames. Images in the video comprise of
face region and other parts of the human body. Frame size
600×800 with 3.16 frames per second processing speed
recorded the highest detection rate which was 95.5% with
one false positive.
2.2 Viola and Jones
Face detection, which aimed to detect human faces from
video sequence using Viola and Jones method is presentedin
[6]. Experimentation on the performance of face detection
was carried out on threedifferentPETSvideodatabasewhich

comprise of face region and other parts of the human body.
The image resolution for the first andsecond video sequence
was 160 ×120 pixels and 720×560 pixels for the third video
sequence. The computation time for the first, second and
third video were 0.637 seconds, 0.371 seconds and 9.258
seconds respectively. Apart from that, this method
successfully detected frontal and slightly rotated-in-plane
faces.
Viola and Jones method also adopted in [7]. The
preprocessing stage was carried out using Viola and Jones
method. First, the image was transformed from Red-Green-
Blue (RGB) colour space to greyscale. Then, histogram
equalization was performed to adjust the contrast. The
output of this stage was used as the candidate image which
may be face or non-face. Experimentation on the
performance of face detection was carried out on 65 images
in CMU database whose sizes ranged between 130×130 to
456×463 pixels. These images consist of face region and
other parts of the human body. When tested, the number of
incorrectly detected faces was reduced from 11 to three.
Besides, detection rate increased from 86.23% to 90.31%.
Apart from that, false positive error ratereducedfrom5.61%
to 1.53%, but false negative error rate remains at 8.16%.
2.3 Haar Features
Real time face detection based on Haar features is
presented in [8]. The paper aimed to detect human faces
rapidly besides achieving a high detection rate. The
advantage was background regions can be quickly discarded
while more computation was done on face-like regions.
Experimentation on the performance of the face detection
was carried out on 1390 video clips which comprise of only
face regions. The detection rate for 24×24 video resolution
which contained various position, scale and orientation of
faces was 89%. For 18×18 and 40×40 video resolution,
detection rate was 85% and 82% respectively.
The paper in [1] aimed to minimize theprocessingtimeof
face detection. Improvised Viola and Jones method for face
detection was based on Haar feature extraction. The
disadvantage of this method was it cannot detect faces with
small sizes and pose variation. Experimentation on the
performance of face detection was carried out with and
without background subtraction on video resolution 160 ×
120 with 2587 frames. Tested imagescompriseoffaceregion
and other parts of the humanbody.Althoughthismethodcan
detect multiple faces, the processing time increased and the
detection time decreased.
Size-based performance analysis of Haar-features for
detection of face images is presented in [9]. Haar features
presented was characterized by their orientation, size and
type. This method was presentedduetoitssimplicityandfast
computation. Experimentation on the performance of face
detection was carried out on 650 images in SCFace database
of size 75×100 and 14×14. The detection rate was higher for
image size ranged from 21×21 to 24×24 pixels. In addition,
image up-scaling resulted in a higher number of false
positives.
In [10], multi-view face detection system which aimed to
minimize search space and boost detection rate is presented.
Face detector presented was able to detect face withroll(-45
to +45) and yaw (-90 to +90) orientations based on Haar
features. Experimentation on the performance of face
detection was carried out on 7500 positive and negative
samples of images which comprise of face region and other
parts of the human body. The detection rate for non-frontal
and frontal faces were 91.30% and 94.10%. True positive,
false negative and false positivevaluesobtainedwere415,25
and 47 for frontal faces and 210, 20 and 42 for non-frontal
faces.
2.4 Other Face Detection Methods
In [11], depth sensor-based face detection for indoor
surveillance is presented. The paper aimed to analyse
complex scene features and integrate the colour and depth
information. The method presented consists of RGBD+ViBe
and 3D Mean-Shift. RGBD+ViBe was used to detect
foreground. Firstly, the CIELAB colour space was applied to
determine the light intensity change. Then, foreground pixel
was calculated. The 3D mean-shift was done to detect facial
features. This stage includes blob analysis, morphological
opening and closing operation and hole filling. Blob analysis
was used to examine the connectivity of similar foreground
pixels. When the blob size was lower than 5,detectedmoving
target was not identified as human. Bounding box width and
height was set to values larger than 64 and 128 respectively.
This box was used as a detection window. Morphological
opening and closingoperationwasadoptedtosuppressnoise
data with a threshold value set to 5. Edge smoothing was
done using a hole filling. This method has a promising
outcome in moving object detection, but it has limitation in
insufficient illumination condition. Experimentation on the
performance of face detection was carried out on 380frames
for bright environment, 403 frames for dark environment
and 250 frames forvariableenvironment.Imagesinthevideo
comprise of face region and other parts of the human body.
When experimented in a bright environment, true positive
rate (TPR) and false positive rate (FPR) were 89% and 3.8%
respectively. As for dark environment, it recorded 88.62%
(TPR) and 5.71%(FPR)whilevariableenvironmentrecorded
90.13% (TPR) and 5.83% (FPR).
The face detection method based on
photoplethysmographyispresentedin[12].Thepaperaimed
to detectfacesbyestimatingphysiologicalmeasures.Thefirst
step involved region of interest (ROI) creation. In the paper,
ROI size used was 40×30 pixelsand 18×18 pixels. Next,heart
rate was estimated by computing the mean pixel values for
RGB colour channel of each ROI. Finally, index of heart rate
was computed to ensure that detected peak matches the real
heart rate. Index ratecloser to oneindicatesahighpossibility
in the presence of living body parts in the ROI.
carried out on video with 640 x 480 pixels stream size which
comprise of only face region. The author claimed that 18×18
pixels detection rate were lower than 40×30 pixels and time
window less than ten seconds also resulted in lower
detection rate.
In [13], video-frame based face detection system which
aimed to detect faces from the video is presented. The
method presented was Libfacedetection which combined
local binary pattern (LBP) feature and Adaboost algorithm.
First, the training samples were labelled and divided into

positive samples and negative samples. For each sample,
average weight distributionwasassigned.Then,theclassifier
was trained to select the bestweakclassifierfromallfeatures
so that the average square error of the sample was
minimized. Finally, the weight of training set was updated.
Face angle of −40° to 40°, −60° to 60° and −90° to 90° can be
detected using this method. Experimentation on the
performance of face detection was carried out on 6100
imagesofCAS-PEAL,FMediaandMaskNetdatasets.Imagesin
the video comprise of only face region. Detection time
obtained for frontal and multiview face were 2.92ms and
3.83ms at 342.5 and 140.4 frames per second respectively.
The face detection method based on multi-scale
histograms is presented in [2]. The paper aimed to improve
computational efficiency. The method presented consists of
preprocessing, coarse-to-fine texture descriptor and
descriptor construction. Preprocessing was done on input
images to normalize its sizeto128×128pixels.Coarse-to-fine
texture descriptor was used to extract feature in images.
Support Vector Machine in the descriptor construction stage
was used as a classifier to distinguish faces and non-faces.
carried out on 4000 positive and negative image samples of
Libor Spacek’s database. Tested imagescompriseofonlyface
region. When experimented, 4 × 4, 8 × 8, 16 × 16 and 32 × 32
block sizes achieved 95% precision for positive samples and
90% for negative samples. Detection time for 4 × 4, 8 × 8, 16
× 16 and 32 × 32 block size were 3.51799ms, 2.63614ms,
1.22268ms and0.833912respectively.Thismethodachieved
10 times faster in detection rate when the size of the block
was set accurately.
Contour-based procedure forface detection and tracking
from video is presented in [14]. The paper aimed to detect
the location of faces from the video. The method presented
was based on moving face contour. First, frames extracted
from video were converted into grayscale images. Then, the
Robert edge detector was used to detect face edges. After
that, Gaussian filtering method was used to remove non-
desired edges and noise followed by computation of logical
operation foredgedetection.Finally,themovingfacecontour
of each frame was determined. This method was used
because the face can be easily detected in separate frame of
two adjacent video frames. Experimentation on the
performance of face detection was carried out on
HONDA/UCSD video database with 640×480 resolution.
Images in the video comprise of only face region. The author
claimedthatcorrectdetectionwasachievedandmovingfaces
were also efficiently tracked.
Human detection using Cascade Head and Shoulder
Detection (CHSD) method is presented in [15]. The paper
aimed to detect face region and filter out non head-shoulder
regions. The method presented consists of initial feature
rejecter, haar-like rejecter and HoG feature classifier. This
method can handle side view detection of ±30° in pan and 0°
to 60° tilt. Experimentation on the performance of face
detection was carried out on 2000 images of the Pascal face
dataset which comprise of face region and other parts of
human body. The presented method achieved 83.9%
detection rate.
3. FINDINGS
Generally, there are three stages which are preprocessing,
feature extraction and classification stage. Based on the past
researches, the most popularmethodsusedforfacedetection
are skin colour segmentation, Viola and Jones method, and
Haar features. Skin colour segmentation method achieved
88% to 98% detection rate. However, detection on images
comprise of only face regionmay result in this high detection
rate. Viola and Jones method, and Haar features achieved
80% to 90% detection rate which was mostly carried out on
images comprised of face region and other parts of the
human body. Not only that, these methods can also detect
slightly rotated faces. Overall, the methods presented was
tested on 60 to 8000 images with 24×24 pixels as the most
common resolution.Sideviewfacedetectionhasbeencarried
out in only four out of 15 research papers.
4. CONCLUSION
In this paper, 15 papers from journals and were reviewed.
Skin colour segmentation achieved 88% to 98% detection
rate , whereas Viola and Jones method, Haar features
achieved 80% to 90% detectionrate.Classificationisthefinal
step in the face detection process which aims to distinguish
faces from non-faces. Themostpopularclassificationmethod
used for skin colour segmentation and Viola and Jones
method is cascade classifier based on Adaboost algorithm.
This classifier discards unnecessary features that is not part
of a face. The other classification method includes back
propagation neural network and support vector machine.
REFERENCES
[1] S. V. Tathe, A. S. Narote, and S. P. Narote, “Human Face
Detection and Recognition in Videos,” Intl. Conference
on Advances in Computing, Communications and
Informatics (ICACCI), Sept. 2016, pp. 2200 – 2205.
[2] C. Y. Lin, J. T. Fu, S. H. Wang, and C. L. Huang, “New Face
Detection Method Based on Multi-Scale Histograms,”
IEEE Second International Conference on Multimedia
Big Data, 2016, p.p. 229 – 232.
[3] U. Priya, S. Vasuhi, and V. Vaidehi, “Face Detection Using
CbCr Color Model in Video,” 3rd International
Conference on Signal Processing, Communication and
Networking (ICSCN), 2015, p.p. 1 – 5.
[4] A. S. Dhavalikar, and R. K. Kulkarni, “Face Detection and
Facial Expression Recognition System,” International
Conference on Electronics and Communication System
(ICECS -2014), 2014, p.p. 1 - 7.
[5] V. Mutneja, and S. Singh, “GPU Accelerated Face
Detection from Low Resolution Surveillance Videos
using Motion and Skin Color Segmentation,” Optik, vol.
157, 2018, p.p. 1155 – 1165.
[6] W. Zou, Y. Lu, M. Chen, and F. Lv, “Rapid Face Detection
in Static Video using Background Subtraction,” 10th

International Conference on Computational Intelligence
and Security, 2014, p.p. 252 – 255.
[7] M. Da’san, A. Alqudah, and O. Debeir, “Face Detection
using Viola and Jones Method and Neural Networks,”
International Conference on Information and
Communication Technology Research (ICTRC2015),
2015, p.p. 40 – 43.
[8] P. I. Rani, and K. Muneeswaran, “Robust Real Time Face
Detection Automatically from Video Sequence Based on
Haar Features,” International Conference on
Communication and Network Technologies (ICCNT),
2014, p.p. 276 – 280.
[9] V. Mutneja, and S. Singh, “Size-based Performance
Analysis of Haar-features for Detection of Facial Images
from Low Resolution SurveillanceVideos,”International
Conference on Intelligent Computing, Instrumentation
and Control Technologies (ICICICT), 2017, p.p. 1191 –
1195.
[10] A. K. Gor, and M. S. Bhatt, “Fast Scale Invariant Multi-
view Face Detection from Color Images using SkinColor
Segmentation & Trained Cascaded Face Detectors,”
International Conference on Advances in Computer
Engineering and Applications(ICACEA), 2015,p.p.688–
694.
[11] T. Hu, H. Zhang, X. Zhu, J. Clunis, and G. Yang, “Depth
Sensor Based Human DetectionforIndoorSurveillance,”
Future Generation Computer Systems,vol. 88,2018,p.p.
540 – 551.
[12] G. Gilbert, D. D’. Alessandro, and F. Lance, “Face
Detection Method Based on Photoplethysmography,”
Workshop on Low-Resolution Face Analysis (LRFA) in
conjunction with 10th IEEE International Conferenceon
Advanced Video and Signal Based Surveillance, 2013,
p.p. 449 – 453.
[13] G. Niu, and Q. Chen, “Learning An Video Frame-Based
Face Detection System for Security Fields,” Journal of
Visual Communication and Image Representation, vol.
55, 2018, p.p.457 – 463.
[14] A. Dey, “A Contour Based Procedure for Face Detection
and Tracking from Video,” 3rd Int'I Conf. on Recent
Advances in Information Technology, 2016, p.p. 83 –
488.
[15] Q. Liu, W. Zhang, H. Li. and K. N. Ngan, “Hybrid Human
Detection and Recognition in Surveillance,”
Neurocomputing, vol. 194, 2016, p.p. 10 – 23.

IRJET- Survey on Face Detection Methods

More Related Content

What's hot (19)

Similar to IRJET- Survey on Face Detection Methods (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Survey on Face Detection Methods