2019 16th International Multi-Conference on Systems, Signals & Devices (SSD'19)

Comparison of Haar-like, HOG and LBP

approaches for face detection in video sequences
Amal Adouani Wiem Mimoun Ben Henia Zied Lachiri
University of Tunis El Manar University of Tunis El Manar University of Tunis El Manar
National Engineering School of, National Engineering School of National Engineering School of
Tunis, Tunis, Tunis,
Signal, Image and Information Signal, Image and Information Signal, Image and Information
Technologies laboratory Technologies laboratory Technologies laboratory
(LR-11-ES17) (LR-11-ES17) (LR-11-ES17)
BP. 37 Belvdre, 1002, Tunis, BP. 37 Belvdre, 1002, Tunis, BP. 37 Belvdre, 1002, Tunis,
Tunisia Tunisia Tunisia
Email: amal.adouani@gmail.com Email: mimoun.wiem@gmail.com Email: zied.lachiri@enit.rnu.tn

Abstract— Face detection is an essential part of any face Several studies have been proposed to compare these
recognition system as a first step to detect faces. This paper approaches. K. Kadir et al. [17] evaluated two face detector
presents a comparative study of three commonly used approaches methods, namely, Haar-like cascade and Linear Binary
for face detection, namely Haar-like cascade, Histogram of Pattern. The authors used image samples from three face
Oriented Gradients with Support Vector Machine and Linear databases: Color FERET database [18], Taarlab database
Binary Pattern cascade. For this aim, video sequences from the [19] and MIT CBCL databases [20]. Their results proved that
Database for Emotion Analysis using Physiological Signals LBP method was faster than her method with 140% and
(DEAP) were explored. The proposed methods were developed detected faces more than Haar with a difference of 4%. J.
using Python language with OpenCV and Dlib libraries. The
Chen et al. [21] provided an improvement of rotated frontal
obtained results show that HOG+SVM approach is more robust
face detection by using the AdaBoost algorithm. They
and accurate than LBP and Haar approaches with an average
detection rate of 92.68%.
applied HOG, Haar-like and LBP features to train it. They
employed various angles and inclinations images from the
Keywords—Face detection; AdaBoost; Haar-like; HOG; MUCT [22] face database during the test. By comparing the
LBP; SVM; video processing; computer vision detection rate, accuracy rate, and false positive rate of the
experimental results of each method, they proved that the
I. INTRODUCTION HOG method was very efficient for detecting inclined faces
Object detection is an interesting computer vision topic with an average detection rate of 47.23%.
that has been studied for years due to the important number S. Guennouni et al. [23] presented a comparative study of
of applications that use it. It is an innovative field that multiple object detection such as face detection, pedestrian
focuses on detecting a specific object such as road signs [1], detection and hand gesture detection. They used both of
people [2], vehicle registration plate [3] and human face that Haar-like cascade and Linear Binary Pattern approaches.
has been gaining much interest [4]. Their goal was to enhance the performance of an embedded
Several facial analysis systems, like facial recognition, system, namely the Texas Instrument DM3730 digital media
facial expression recognition, and facial tracking systems, processor. They used low resources to achieve a real-time
can be performed only when a face is well detected in a detection. Haar-like cascade method achieved the highest
given image or video. However, face detection is considered results with an accuracy rate of 96.24% compared to 94.75%
as an extremely difficult research task due to many for LBP. Y. Kortli et al. [24] tested numerous face
challenges such as pose variation, different locations, descriptors, Local Binary Patterns Histogram (LBPH), Local
illumination, expression variation, etc. Binary Patterns (LBP) and Histograms of Oriented Gradients
To resolve these problems, several approaches have been (HOG) using the standard FERET data set. They classified
proposed. Viola and Jones [5] introduced one of the first all of the methods using the SVM classifier. The LBPH
classical face detection methods [6] [7] [8]. Their paper technique outperformed the others in terms of accuracy and
contained 4 stages: (1) Haar like features to detect features, efficiency of face detection. It has achieved an average
(2) Integral Image to quickly calculate them, (3) Adaboost to detection rate greater than 98.04%.
select potential features and (4) Classifier Cascade to rapidly Many of these works aimed to improve and compare the
discard non faces. face detection algorithms using images rather than real-time
The Histogram of Oriented Gradients method has been performance evaluation. Consequently, the obtained results
widely used for face detection [9] [10] [11]. Dalal et al. [12] could be severely misinterpreted. M. Sheng et al. [25]
presented their supplementary work on pedestrian detection focused on the lack of a real time face detection method.
that is based on the idea that the edge directions or the Hence, they tried to improve face detection in video
intensity of the gradient distribution contains important sequences using complex background. Their proposed
information. Consequently, the shape and the appearance of HAAR method achieved an error rate inferior to 5%.
a local object can be easily defined. W. Zou et al. [26] proposed a new method capable of
Ojala et al [13] proposed an efficient type of visual using the Viola-Jones face detector to rapidly detect faces in
descriptor for describing ordinary textures and shape of a static video. This method provided promising results by
digital image called Local Binary Pattern (LBP). Since, achieving a better computation time and less false positive
many extensions and improvements have been introduced to than the Viola Jones method. In addition, T. Malach et al.
the LBP method [14] [15] [16]. [27] attempted to evaluate the Viola-Jones face detector
algorithm by testing it on different databases that could

978-1-7281-1820-8/19/$31.00 ©2019 IEEE 266

 2019 16th International Multi-Conference on Systems, Signals & Devices (SSD'19)

represent real time scenarios and their obtained results A. Affective Multimodal Database DEAP
achieved a 78.8% True Positive rate. Various databases are established for the face detection
This paper aims to compare three of the most commonly task to test the system’s performance, such as JAFFE [29],
used approaches for face detection namely Haar-like YALE [30], CMU PIE [31] and FERET [18]. The frontal
cascade, Histogram of Oriented Gradients with Support face video sequences are selected to compare these face
Vector Machine and Linear Binary Pattern cascade detection methods instead of static face images. Thus, the
algorithms. The tests were applied using the affective DEAP multimodal database [28] based on analyzing elicited
multimodal DEAP database [28] that contains frontal face’s human emotions is selected. The DEAP frontal face videos
video sequences. Three sections are organized as follows: were recorded from 22 participants watching 40 music
First section presents details on each of the face detection affective videos. Only 19 subject data were publicly
methods. Second section presents and discusses the obtained available. In this paper, only 7 participants whom are used
results. The third section contains the conclusion and future and each session contains almost 40 videos.
works.
B. Haar-like Features
II. PROPOSED APPROACH METHODS The Haar-like features algorithm consists of training a
In this work, three different steps are realized to compare strong classifier using a big number of examples. It aims to
the discussed face detection methods: (1) the video sequence detect the possible presence of the face in any image by
is selected from the affective multimodal DEAP database, exhaustively traversing it at all positions and in all possible
(2) the face detection algorithm is applied to this video and sizes.
(3) the algorithms are evaluated and compared. This proposed algorithm has four stages:
The flowchart of this study is presented in figure1. 1) Haar Features
Viola-Jones algorithm uses rectangle filters, called Haar-like
DEAP database
(278 video sequences) features rather than working directly on pixel values and to
be more efficient and faster. These filters are used to test the
presence or not of the face. There exist three types of Haar-
Input video sequence like features according to [5]: (1) the two upper squares, (2)
the line feature and (3) the center-surround feature (figure
2).
Convert video to grayscale

Apply face detection technique

Face detection

HAAR LBP HOG

Classification
Fig. 2. Example of Haar-like features

2) Integral Image
Face Not face
Integral image is used to quickly and efficiently calculate
any sum of pixel values in constant time and at a very low
computational cost [5]. This technique consists of creating
Frames extraction and registration
an image with the same size as the original image, called a
supplementary Area Table. The pixels left and above of any
Techniques comparison

pixel (x, y) in the original image are summed up (equation

(1)). Their results present the value of this pixel in the new
Frames evaluation
table.

𝑖𝑖𝑖𝑖(𝑥𝑥, 𝑦𝑦) = ∑𝑥𝑥 ′ ≤𝑥𝑥,𝑦𝑦 ′ ≤𝑦𝑦 𝑖𝑖(𝑥𝑥 ′ , 𝑦𝑦 ′ ) (1)

Face correctly Face falsely detected
detected or not detected Where ii(x, y) presents the integral image pixel values and
i(x, y) presents the original image pixel values of point (x,y).
Rectangle features can be rapidly computed after using the
Metrics calculation and approaches integral image. Calculating the sum of pixel values within a
comparison rectangular area can be done easily by accessing only to four
values of the integral image instead of summing up all the
pixels. An example is shown in figure 3. If A, B, C, D are the
Fig. 1. Proposed comparative study process corners values of the summed table , then the sum within this
rectangle can be computed using the equation (2).
𝑠𝑠𝑠𝑠𝑠𝑠 = 𝐷𝐷 + 𝐴𝐴 − 𝐵𝐵 − 𝐶𝐶 (2)

978-1-7281-1820-8/19/$31.00 ©2019 IEEE 267

 2019 16th International Multi-Conference on Systems, Signals & Devices (SSD'19)

gradient feature vectors are computed for each cell ( see

figure 5). These gradient feature vectors are then added to
A B obtain the feature vector for a single frame. Finally, all
C D gradient feature vectors extracted from different images are
(x,y)
concatenated to obtain one long vector which is the HOG
feature vector. The later will be used as input to SVM
classifier.
Fig. 3. Computing Sum within a given Rectangle Blocks
3) Adaboost:
A big number of possible features are collected from every
sub window. Hence, AdaBoost, a popular boosting method
is used to exclude the irrelevant features. It selects only the
most informative ones by combining multiple “weak
classifiers” to build a single “strong classifier”. Cells
4) Cascading classifiers
Cascading classifiers are used to quickly discard non-face
and avoid any unnecessary processing. In cascading, a Histogram
cascade classifier is composed of many stages that each
contains a strong trained classifier. If a given sub window
fails at any stage, it will be immediately discarded.
Adaboost is used to design a cascade by preparing every
single stage. It combines classifiers together to construct the Fig. 5. Hog feature extraction
stages. Figure 4 shows an example of the cascade
D. Linear Binary Pattern
classification process that contains 4 stage: (1), (2), (3) and
(4). Linear Binary Pattern (LBP) has a very high performance
in many different applications such as image retrieval [32],
surface inspection [33], texture classification [34],
All segmentation [35] and face recognition [36]. The biggest
sub-windows
advantages of the LBP features are their illumination
invariance and their robustness. This approach consists of
Further dividing the image into several blocks and finally extracting
1 2 3 4 processing and concatenating the different features into a feature vector.
This vector will be used later as a facial descriptor. The basic
Rejected sub-windows
methodology of LBP consists of dividing the input image
into non overlapping small local regions called cells. This
Fig. 4. Cascade classification Schematic descriptor compares the central pixel to each one of its 8
neighbor pixels. If the value of the central pixel is greater
C. Histogram of Oriented Gradients than or equal to the neighbor value, then 1 is written,
The Histogram of Oriented Gradients (HOG) method otherwise, 0 is written. Next, a binary number is obtained by
allows the histogram to be easily calculated. In fact, every reading the updated pixel values in clockwise. Then, it is
key point owns a generated HOG feature. The neighboring converted into a decimal number. This number is the new
area of each one is divided into small blocks called cells. A value of the central pixel (figure 6).
local 1-D histogram of gradient directions is compiled for the
pixels within each cell. Consequently, the descriptor is the
combination of all these histograms. The gradient vector is
calculated as follows: 58 51 48 1 1 1
𝐺𝐺𝑥𝑥 (𝑥𝑥, 𝑦𝑦) = 𝐻𝐻(𝑥𝑥 + 1, 𝑦𝑦) − 𝐻𝐻(𝑥𝑥 − 1, 𝑦𝑦) (3)
80 16 15 1 1 255
𝐺𝐺𝑦𝑦 (𝑥𝑥, 𝑦𝑦) = 𝐻𝐻(𝑥𝑥, 𝑦𝑦 + 1) − 𝐻𝐻(𝑥𝑥, 𝑦𝑦 − 1) (4)
The horizontal gradient of the image pixel Gx(x, y) and the 10 2 9 0 0 0
vertical gradient Gy(x, y) are presented in equation (3) and
(4). Fig. 6. LBP conversion to binary

The gradient amplitude and direction of the pixel (x,y) are The LBP descriptor is calculated on a defined region as
respectively presented in equation (5) and (6). follows:
𝐺𝐺(𝑥𝑥, 𝑦𝑦) = �𝐺𝐺𝑥𝑥 (𝑥𝑥, 𝑦𝑦)2 + 𝐺𝐺𝑦𝑦 (𝑥𝑥, 𝑦𝑦)2 (5) 𝐿𝐿𝐿𝐿𝐿𝐿(𝑥𝑥𝑐𝑐 , 𝑦𝑦𝑐𝑐 ) = ∑7𝑛𝑛=0 𝑠𝑠 (𝑙𝑙𝑛𝑛 − 𝑙𝑙𝑐𝑐 )2𝑛𝑛 (7)
Where (xc, yc) are the coordinates of the central pixel, lc the
−1 𝐺𝐺𝑥𝑥 (𝑥𝑥,𝑦𝑦) gray value of the central pixel, ln the gray values of the
𝑎𝑎(𝑥𝑥, 𝑦𝑦) = tan ( ) (6)
𝐺𝐺𝑦𝑦 (𝑥𝑥,𝑦𝑦) neighborhood pixels and s is defined as follows:
The given frame is divided into either rectangular or circular 1 𝑘𝑘 ≥ 0
𝑠𝑠(𝑘𝑘) = � (8)
regions of size N*N pixels, called cells. Furthermore, 0 𝑘𝑘 ≤ 0

978-1-7281-1820-8/19/$31.00 ©2019 IEEE 268

 2019 16th International Multi-Conference on Systems, Signals & Devices (SSD'19)

LBP was extended to be able to use the neighborhoods of

various sizes [36]. The idea was to use a bilinear
interpolation and a circular neighborhood of the pixel values
to handle any variable radius. The notation (P, R) was used
for pixel neighborhoods where P is the sampling points and
R is the radius. Uniform patterns technique was another
extension of the original LBP operator. It was used to
decrease the number of values from 256 to 59. In fact, it has
been proved that some binary patterns contain more
important information than others. The texture primitives
obtained by these bins describe different types of areas such
as edge, line end, corner, etc. The occurrences of the uniform
patterns in an image are collected into LBP histogram. Then,
the extracted LBP histograms are concatenated to finally
form an enhanced spatial histogram called feature histogram.
The LBP feature vector is processed to classify the images
by using extreme learning machines such as SVM and
cascade classifiers. Fig. 7. Sample results taken from DEAP database

III. RESULTS AND DISCUSSIONS Table I illustrates the obtained results of using the Haar-like
cascade detector. It proves that this approach is very accurate
These methods were implemented using the python while detecting face region by detecting the whole face at
language with OpenCV image processing library [37] and different scales. It achieves 99.7% detection rate. However, it
dlib library [38]. In this paper, two metrics are considered: doesn’t perform if the intensity between the dark and bright
the True Positive Rate (TPR) and the False Negative Rate area is very low or if the face is oriented: Even a small
(FNR). True Positive Rate measured the proportion of faces change in the orientation of the face is enough to not detect
that were correctly identified as such. It was calculated using it. This constatation can be clearly seen through the results of
equation (9). participants 4 and 5 where the detection rate didn’t reach
𝑇𝑇𝑇𝑇 35%: The face of participant 4 was a little inclined during the
𝑇𝑇𝑇𝑇𝑇𝑇 = (9) whole test and as for participant 5, his gray scale face frame
(𝑇𝑇𝑇𝑇+𝐹𝐹𝐹𝐹)
was a bit bright. So, the black features were not detected.
Where TP is the true positive frames number and FN is the
true negative frames number. TABLE I. HAAR-LIKE CASCADE RESULTS
False negative Rate measured the proportion of faces which
yielded negative test outcomes with the test using the METRICS
following equation. It is calculated as follows: Detected
Detected frames TPR(%) FNR(%)
𝐹𝐹𝐹𝐹 faces
𝐹𝐹𝐹𝐹𝐹𝐹 = (10) Participant 1 120012/120000 118796 98.99 48.99
(𝑇𝑇𝑇𝑇+𝐹𝐹𝐹𝐹)

The test results were computed using the following steps: Participant 2 119984/120000 119984 99.98 0.3
First, the public multimodal DEAP database was chosen for Participant 3 119977/117000 116968 99.97 21.7
testing the ability of the face detection approaches. 7 sessions Participant 4 41564/120000 41564 34.63 65.75
were selected out of 19 sessions. Each session contained Participant 5 22814/117000 22814 19.49 92.51
almost 40 frontal face video sequences of a given
Participant 6 115572/120000 115572 96.31 11.17
participants while he was watching affective music video.
Overall, 278 frontal videos sequences from the DEAP Participant 7 117749/120000 116753 97.29 0.69
database were used during the experimental tests. Each
video sequence contained 3000 frame images. A total of Table II resumes the obtained results of the Hog detector.
834000 frame images were extracted and classified of each This method has a very high True Positive Rate that reaches
face detection method. First, the Haar-like cascade detector 99%. It detects both of frontal faces and slightly the non
was applied on each given video. While detecting faces, the frontal ones with a small occlusion because it is based on
numbers of the detected frames were calculated and the 3000 shape detection. However, if the shape is not clear or the face
frame images were extracted at the same time. Later, the is much rotated, detection cannot be achieved. Consequently,
extracted frame images were classified manually to compute participant 4 achieved a low detection rate where half of the
the false detected faces. Then, the other metrics were face frame images were not detected. This was mainly due to
computed including the average metrics values of all the the lack sharp feature especially around the eyes, nose bridge
videos in each. The same process was applied to test the and the mouth and due to some occlusion where the subject
tried to fix it glasses several times.
performances of LBP and HOG methods. Figure 7 shows 4
examples of the extracted frame images from DEAP video
sequences of the DEAP database.

978-1-7281-1820-8/19/$31.00 ©2019 IEEE 269

 2019 16th International Multi-Conference on Systems, Signals & Devices (SSD'19)

TABLE II. HOG RESULTS As mentioned previously, this paper deals with frontal
face detection using video sequences. Consequently and on
METRICS
future work, more face detection challenge will be
Detected frames Detected faces TPR(%) FNR(%) considered such as orientation variance, illumination, poses
Participant 1 119986/120000 119986 99.98 0.02 and facial expression changes by testing other databases such
Participant 2 120000/120000 120000 100 0
as RECOLA, JAAFE and FRET databases. Moreover, to
improve this work, other face detection methods can be
Participant 3 117000/117000 117000 100 0
implemented and evaluated. We mention the Skin color
Participant 4 64721/120000 64721 53.93 46.07 method, the SIFT feature descriptor and the artificial
Participant 5 117000/117000 117000 92.68 7.32 intelligence techniques (Fuzzy Logic, Artificial Neural
Participant 6 119880/120000 119880 99.99 0.01 Network, Deep Learning ...).
Participant 7 117367/120000 117367 97.74 2.26
ACKNOWLEDGMENT
The authors of this paper would like to express their
Table III presents the results of the LBP detector. It proves sincere thanks to the DEAP database teams for providing this
that the obtained results are not very promising. LBP
affective multimodal database to test this work:
achieves the highest False Negative Rate of 92.51. The
http://www.eecs.qmul.ac.uk/mmv/datasets/deap/index.html
participants 1, 4 and 5 presented the worst detection rates
that are inferior to 52% since LBP could not capture well REFERENCES
their local patterns.
[1] W. Farhat, H. Faiedh, C. Souani, and K. Besbes, “Embedded system
for road sign detection using MicroBlaze,” in 2015 IEEE 12th
TABLE III. LBP RESULTS
International Multi-Conference on Systems, Signals Devices
(SSD15), 2015, pp. 1–5.
METRICS
[2] T. Wang, C. Chang, and Y. Wu, “Template-based people detection
Detected frames Detected faces TPR(%) FNR(%) using a single downward-viewing fisheye camera,” in 2017
International Symposium on Intelligent Signal Processing and
Participant 1 61353/120000 61216 51.01 48.99 Communication Systems (ISPACS), 2017, pp. 719–723.
Participant 2 11967/120000 119651 99.7 0.3 [3] S. Issaoui, R. Ejbeli, T. Frikha, and M. Abid, “Embedded approach
Participant 3 91658/117000 91658 78.3 21.7 for edge recognition: Case study: Vehicle registration plate
Participant 4 4111/120000 4111 34.25 65.75 recognition,” in 2016 13th International Multi-Conference on
Systems, Signals Devices (SSD), 2016, pp. 336–341.
Participant 5 929/117000 929 7.49 92.51
[4] C. Maaoui, F. Abdat, and A. Pruski, “Wavelet transform based facial
Participant 6 106651/120000 106607 88.83 11.17
feature points detection,” in 2017 14th International Multi-Conference
Participant 7 119180/120000 119178 99.31 0.69 on Systems, Signals Devices (SSD), 2017, pp. 5–10.
[5] P. Viola and M. Jones, “Rapid object detection using a boosted
Table IV compares the obtained results of each method. cascade of simple features,” in Proceedings of the IEEE Computer
It clearly shows that the Histogram of Oriented Gradients Society Conference on Computer Vision and Pattern Recognition,
(HOG) is the most accurate out of the three methods by 2001, pp. 511–518.
achieving the highest performance. It detects 14.45% more [6] S. Ma and L. Bai, “A face detection algorithm based on Adaboost and
new Haar-Like feature,” in 2016 7th IEEE International Conference
faces than the Haar-like cascade and 32.31% more faces than on Software Engineering and Service Science (ICSESS), 2016, pp.
the Linear Binary Pattern cascade (LBP). 651–654.
[7] M. A. S. Mollah, M. A. A. Akash, M. Ahmed, and M. A. H. Akhand,
TABLE IV. COMPARISON OF FACE DETECTION METHODS “Improvement of haar feature based face detection incorporating
human skin color analysis,” in 2016 International Conference on
Methods Medical Engineering, Health Informatics and Technology (MediTec),
HAAR HOG LBP 2016, pp. 1–6.
Detected frames 653484 772954 503516 [8] L. Cuimei, Q. Zhiliang, J. Nan, and W. Jianhua, “Human face
Metrics

Detected faces 652451 772954 503350 detection algorithm via Haar cascade classifier combined with three
additional classifiers,” in 2017 13th IEEE International Conference on
TPR(%) 78.23 92.68 60.37
Electronic Measurement Instruments (ICEMI), 2017, pp. 483–487.
FNR(%) 21.76 7.31 39.64 [9] N. Dalal and B. Triggs, “Histograms of Oriented Gradients for
Human Detection” in 2005 IEEE Computer Society Conference on
IV. CONCLUSION Computer Vision and Pattern Recognition (CVPR’05), 2005, pp.
In this paper, three face detectors were reviewed 886–893.
and evaluated: (1) Haar-like cascade, (2) Histogram of [10] A. Dehghani, D. Moloney, and X. Xu, “Face detection speed
Oriented Gradients (HOG) with Support Vector Machine and improvement using bitmap-based Histogram of Oriented gradien,” in
2017 International Conference on Systems, Signals and Image
(3) Linear Binary Pattern cascade (LBP). Their performances Processing (IWSSIP), 2017, pp. 1–5.
were compared based on the False Negative Rate (FNR) and
[11] Y. Kim, H. Shahdoost, S. Jadhav, and C. S. Gloster, “Improving the
the True Positive Rate (TPR). Frontal face video sequences Accuracy of Arctan for Face Detection,” in 2017 IEEE 25th Annual
from the free multimodal DEAP were used during the test. International Symposium on Field-Programmable Custom Computing
This comparative study proved that the hog face detector was Machines (FCCM), 2017, pp. 202–202.
the most accurate out of the three methods. It achieved the [12] D. Huang, C. Chen, T. Chen, J. Wu, and C. Ko, “Real-Time Face
highest detection rate (92.68%). The Haar face detector Detection Using a Moving Camera,” in 2018 32nd International
achieved the second-best detection rate almost 14% inferior Conference on Advanced Information Networking and Applications
Workshops (WAINA), 2018, pp. 609–614.
to the HOG face detector. The LBP face detector achieved
[13] T. Ojala, M. Pietikäinen, and T. Mäenpää, “Multiresolution gray-scale
the worst detection rate which was 32% inferior to the HOG and rotation invariant texture classification with localbinary patterns,”
face detector. This approach was very fast and not accurate. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.
It detected many false positives during the test. 24, no. 7, pp. 971–987, 2002

978-1-7281-1820-8/19/$31.00 ©2019 IEEE 270

 2019 16th International Multi-Conference on Systems, Signals & Devices (SSD'19)

[14] L. Acasandrei and A. Barriga, “Embedded Face Detection [26] W. Zou, Y. Lu, M. Chen, and F. Lv, “Rapid Face Detection in Static
Application Based on Local Binary Patterns,” in 2014 IEEE Intl Conf Video Using Background Subtraction,” in 2014 Tenth International
on High Performance Computing and Communications, 2014 IEEE Conference on Computational Intelligence and Security, 2014, pp.
6th Intl Symp on Cyberspace Safety and Security, 2014 IEEE 11th 252–255.
Intl Conf on Embedded Software and Syst (HPCC,CSS,ICESS), [27] T. Malach, P. Bambuch, and J. Malach, “Face detection in video
2014, pp. 641–644. sequences,” in Proceedings of 22nd International Conference
[15] Z. Jun, H. Jizhao, T. Zhenglan, and W. Feng, “Face detection based Radioelektronika 2012, pp. 1–4.
on LBP,” in 2017 13th IEEE International Conference on Electronic [28] S. Koelstra, C. Muhl, M. Soleymani, JS. Lee, A. Yazdani, T.
Measurement Instruments (ICEMI), 2017, pp. 421–425. Ebrahimi, T. Pun, A. Nijholt, I. Patras, “DEAP: A Database for
[16] W. H. Alobaidi, I. T. Aziz, T. Jawad, F. M. F. Flaih, and A. T. Azeez, Emotion Analysis Using Physiological Signals,” IEEE Transactions
“Face detection based on probability of amplitude distribution of local on Affective Computing, vol. 3, no. 1, pp. 18-31.
binary patterns algorithm,” in 2018 6th International Symposium on [29] J. Michael, A. Shigeru, K. Miyuki, and G. Jiro, “Coding facial
Digital Forensic and Security (ISDFS), 2018, pp. 1–5. expressions with gabor wavelets,” in IEEE Intl. Conf. on Automatic
[17] K. Kadir, M. K. Kamaruddin, H. Nasir, S. I. Safie, and Z. A. K. Bakti, Face andGesture Recognition, 1998, pp. 200 – 205.
“A comparative study between LBP and Haar-like features for Face [30] Georghiades, A.S. and Belhumeur, P.N. and Kriegman, D.J., "From
Detection using OpenCV,” in 2014 4th International Conference on Few to Many: Illumination Cone Models for Face Recognition under
Engineering Technology and Technopreneuship (ICE2T), 2014, pp. Variable Lighting and Pose", IEEE Trans. Pattern Anal. Mach.
335–339. Intelligence, 2001, pp. 643-660.
[18] P. J. Phillips, H. Moon, S. A. Rizvi, and P. J. Rauss, “The FERET [31] T. Sim, S. Baker, and M. Bsat, “The CMU Pose, Illumination, and
evaluation methodology for face-recognition algorithms,” IEEE Expression (PIE) database,” in Proceedings of Fifth IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 22, International Conference on Automatic Face Gesture Recognition,
no. 10, 2000,pp. 1090–1104. 2002, pp. 53–58.
[19] 2004, “Stimulus images courtesy” of M. J. Tarr, Center for the [32] H. Singh and D. Agrawal, “An analysis based on local binary pattern
NeuralBasis of Cognition and Department of Psychology, Carnegie (LBP) and color moment (CM) for efficient image retrieval,” in 2016
MellonUniversity (Brown University), webpage on The CNBC Wiki International Conference on Emerging Technological Trends
[Online].Available:http://cbcl.mit.edu/softwaredatasets/heisele/facere (ICETT), 2016, pp. 1–6.
cognitiondatabase. html/.
[33] M. Quintana, J. Torres, and J. M. Menéndez, “A Simplified Computer
[20] B. Weyrauch, J. Huang, B. Heisele & V. Blanz, “Component-based Vision System for Road Surface Inspection and Maintenance,” IEEE
face recognition with 3D morphable models,” in: First IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 3, pp.
Workshop on Face Processing in Video, Washington, DC, 2004. 608–619, Mar. 2016.
[21] J. Chen, I. Tang, and C. Chang, “Enhancing the Detection Rate of [34] A. R. Backes and J. J. de M. Sá, “Texture Classification Using Fractal
Inclined Faces,” in 2015 IEEE Trustcom/BigDataSE/ISPA, 2015, vol. Dimension Improved by Local Binary Patterns,” in 2018 26th
2, pp. 143–146. European Signal Processing Conference (EUSIPCO), 2018, pp. 1312–
[22] S. Milborrow, J. Morkel, and F. Nicolls, “The muct landmarked face 1316.
database,” Journal of Pattern Recognition Association of South [35] H. Du, Z. Wang, D. Wang, and X. Wang, “Multi-Feature Fusion
Africa, 2010. Method Applied in Texture Image Segmentation,” in 2018 14th
[23] S. Guennouni, A. Ahaitouf, and A. Mansouri, “A Comparative Study International Conference on Computational Intelligence and Security
of Multiple Object Detection Using Haar-like Feature Selection and (CIS), 2018, pp. 135–139.
Local Binary Patterns in Several Platforms,” Model. Simul. Eng., vol. [36] M. D. C, K. B. Raja, and V. K. R, “Spatial Domain Face Recognition
2015, pp. 1-8. System Using Convolution of PDV and LBP,” in 2018 International
[24] Y. Kortli, M. Jridi, A. A. Falou, and M. Atri, “A novel face detection Conference on Intelligent Informatics and Biomedical Sciences
approach using local binary pattern histogram and support vector (ICIIBMS), 2018, vol. 3, pp. 321–326.
machine,” in 2018 International Conference on Advanced Systems [36] A. Hadid, G. Zhao, T. Ahonen, and M. Pietikainen, Face analysis
and Electric Technologies (IC_ASET), 2018, pp. 28–33.
using Local binary Patterns, World Scientific Volume, 2008, pp.347-
[25] M. Sheng and Z. Lan, “Face detection in video sequence with
350.
complex background,” in 2008 IEEE International Conference on
Mechatronics and Automation, 2008, pp. 432–437. [37] G. Bradski and A. Kaehler, Learning OpenCV: Computer vision with
the OpenCV library. ” O’Reilly Media, Inc.”, 2008.
[38] “dlib C++ Library.” [Online]. Available: http://dlib.net/.

978-1-7281-1820-8/19/$31.00 ©2019 IEEE 271