ISSN (Print) : 2319-8613

ISSN (Online) : 0975-4024 Dinesh Kunar P et al. / International Journal of Engineering and Technology (IJET)

Improved Edge Detection based Fast Face

Detection Method using Enhanced Fourier
Transform (IED-FFD) towards Facial
Expression Recongnition
DINESH KUNAR P
Ph.D. Research Scholar, PG and Research Department of Computer Science
Dr. N.G.P. Arts and Science College, Coimbatore

Dr. B. Rosiline Jeetha

Professor and Head, PG and Research Department of Computer Science
Dr. N.G.P. Arts and Science College, Coimbatore
Abstract - Facial expression recognition extends in various real world scenario that includes safety,
scientific and many commercial applications. Face detection through edge detection is focussed in this
research before recognizing facial expressions. Hence, in this paper, an improved sobel operator is
employed. The improved sobel operator makes use of geodesic distance. To identify the position of edge,
the minimum distance between the image and the elements of features with the help of geodesic distance
are used. After that, an enhanced fast fourier transform (E-FFT) is employed for face detection.
Performance metrics such as time taken for edge detection, time taken for face detection and accuracy
are taken into account. From the results it is evident that the proposed work consumes less time for edge
detection and face detection. Also it is noteworthy that the accuracy is improved significantly.
Keywords: Facial expression recognition, edge detection, face detection, fast fourier transform, accuracy.
1. INTRODUCTION
Face is a vital part is social communications. Ekman and Frisen [18] mentioned that, happy, sad, anger, fear,
disgust and surprise are the six basic expressions that are freely documented across diverse cultures. A system
developed for contemplating facial reactions mechanically through a human-computer interaction, is known as
Facial expression recognition. Edges are basically the noticeable variation of intensities in an image. Edges help
to identify the location of an object and the boundary of a particular entity in the image. It also helps in feature
extraction and pattern recognition. Hence, edge detection is of great importance in computer vision. So far, most
of the researchers have chosen software for implementation of basic edge detection algorithms and their
variations [19]. But, it has been established that it is not an efficient approach for real time applications. Face
detection helps in FER method for faster recognition of human expressions. This research article proposes
Improved Edge Detection based Fast Face Detection Method using Enhanced Fourier Transform (IED-FFD)
towards Facial Expression Recongnition. This paper is organized as follows. Section 2 gives a brief review of
literature. Section 3 describes the proposed work. Section 4 provides results and discussions. Section 5 offers
concluding remarks.
2. RELATED WORKS
Amer et al. [1] compared all the edge detection algorithms to study their performances. The results showed that
Sobel and Canny edge detectors are not as much of sensitive to random noise in an image as compared to
Robert’s and Prewitt operators. Cui et al. [2] deliberated the recompenses and difficulties of edge detection
algorithms and matched them. The evaluationindicates that binary morphology operator attains better results. A
system namely bordering close has also been used so as to achieve clear and integral image profile. [3] Deng et
al. developed an algorithm in which fusion technology is used. In this an improved Sobel operator is introduced
which is combined with wavelet transform, canny and prewitt operators. The algorithm effectively improved the
results of edge detection. Chinu et al.[4] discuss a hybrid strategy for color based image edge detection. This
approach has been mainly implemented for overcoming poor edge localization, high sensitivity to noise, edge
detection of images with complex background and detecting color edges. Both sequential and parallel
approaches have been used and compared which depicted that parallel strategy shows a performance gain of
68%. To measure the accuracy of edge detection various metrics are used like PSNR, speed up, efficiency etc.
[5]. An algorithm is developed in which improved canny operator is used. In this improved operator,
morphological filter replaces the Gaussian filter. The morphological filtering pre-treats the noise of image,
keeps the edge strength and details, thus achieving better accuracy. Gao et al.[6] proposed a method in which
Sobel edge detection is combined with soft-threshold wavelet de-noising on images which have White Gaussian

DOI: 10.21817/ijet/2017/v9i6/170906141 Vol 9 No 6 Dec 2017-Jan 2018 4464

ISSN (Print) : 2319-8613
ISSN (Online) : 0975-4024 Dinesh Kunar P et al. / International Journal of Engineering and Technology (IJET)

noises. Firstly,
F soft-thhreshold waveelet denoising is used to reemove the noiises from the image and th hen Sobel
operator is applied to it. i The results show an imprrovement in th he accuracy off edge detectioon.
As far asa related worrks in face detection
d is concerned, reccently, many appearance bbased face reecognition
systems are
a developedd. Such system ms aimsto lessen the comp putation time thatt is requireed for detectio
on of face
images [113]. The appeearance based face recognitiion systems use face imagee pixel intensitty values direcctly as the
feature values for recoognition and are a representedd using singlee valued variabble. But such single valued d variables
may not be able to cappture the variation of featuure values of the t images off the same subbject and will also have
high dimmensional featuures data. One of the technniques adapted d by researchhers to developp efficient sysstem with
reduced number
n featurres is method-- ology of feaature learning.. Feature learnning or active learning play ys a major
role to diiscriminately represent or construct
c trainning image daata (extracted feature) to immprove the perrformance
of classiffication task [7–10].
[ Hencee, recently a number
n of acttive learning methods havee been propossed which
combine representivenness and inforrmativeness of o training sammples to attaiin better classsification accuuracy [8].
Chong Peng
P et al. propose
p a suppervised featuure learning model for cllassification oof both low and high
dimensioonal data [11].. The techniquue estimates reegression vecttor using a diss- criminative regression ap pproach to
estimate the similaritty between the t test data and training g data. Howeever, classificcation task be- b comes
challengiing, when low w or high dim mensional imaage data are affected
a by diistortions suchh as illuminattion, pose
variationn, presence off occlusion annd geometricaal variations etc. e [12]. To handle
h such a class of disttortions, a
mathemaatical framewoork for multi- class classificcation of high dimensional image data haas been pro- posed. p The
minimax framework estimatese an optimal repreesentation mo odel that miniimizes the fittting error un nder listed
distortionns to the dataa in an appliccation of interest. Further, categorical information
i iss derived based on the
estimatedd model and reegression moddels used for classification
c [12].
During thhe contemporrary years sevveral research works are caarried out as far f as face reccognition is concerned.
c
Many researchers aim m to reduce thet time comp mplexity in terrms of compuutation time ttaken for faccial image
detectionn. It is to be noted
n that therre are face reecognition sysstems that makke use of face image pixell intensity
values diirectly as the feature valuees for recogniition. Such sy ystems / methhods the repreesentation maade use of
single vaalued variable. In some literrature, it is mentioned that single valuedd variables aree not capable enough e to
seizure thhe variation off feature valuees of the faciaal images that has high dimeensional featuures.
3. PR
ROPOSED WORK W
3.1. Improved Sobel Operator
O forr Edge Detecttion in Facial Expression Recognition
R IImages
The sobeel operator is a famous edgge detection mechanismm nd can be appplied to wide variety of im
an mages that
includes digital color images
i and grrayscale imagges. In this secction, a brief introduction
i aabout sobel op
perator for
edge deteection is preseented and thenn the improvem ment work is presented
p lateer. The sobel ooperator has a couple of
3×3 convvolution kerneels and is depicted in Fig.1.

Fig.1. Convoolution kernel in Sobel

S operator

The Sobeel operator’s kernel

k structurre helps to rettort in the best way to edgees running verrtically and ho
orizontally
based onn the pixel griid, one kernel for each of the two perp pendicular orieentations. Theese kernels arre applied
distinctlyy to the input facial
f expression image in order
o to obtain
n discrete meaasurements off the incline co
omponent
in each orientation
o andd represented as Gx and Gy. In this work k, these are em
mbeeded togetther for discov vering the
completee magnitude of o the incline at each pointt and the orieentation of thaat incline. Thhe incline mag gnitude is
mathemaatically represeented as

G G
Gx 2  Gy 2 … (1)
Inorder too quickly proccess, the equaation (1) is moodified as
G  Gx  G y … (2)

DOI: 10.21817/ijet/2017/v9i6/170906141 Vol 9 No 6 Dec 2017-Jan 2018 4465

ISSN (Print) : 2319-8613
ISSN (Online) : 0975-4024 Dinesh Kunar P et al. / International Journal of Engineering and Technology (IJET)

The angle of direction of the edge for spatial incline is represented as,
 Gy 
  arctan   … (3)
 Gx 
It is considered that giis the ith image where i [0 ,1] . In order to identify the position of edge, the minimum
distance between the image and the elements of features are used and is given as,
d *pos  arg min p (d pos | dist ( Fn ,  img ) … (4)

In this work, the geodesic distance is used. In the input facial expression image I (x ) , the geodesic distance
D( xi , x j ) between the pixels xi and x j is defined as the shortest path that connects xi and x j . It is
mathematically represented as,

D xi , x j 
min l ( )

  Pxi , x j 
0
1  2 (I ( s ).' ( s )) 2 ds … (5)

In the equation (5), Pxi , x j is the set of all available paths between xi , x j ,  is the parameterized depiction by
arclength on the path, I (s ) is the incline value,  ' ( s ) is the unit vector tangent to the direction of the
path. The factor  weights the influence of the image incline versus spatial distance. The input facial expression
image is the main processing target, and the geodesic distance is computed only by accumulating neighboring
pixel values as in equation.6.
n n
d ()   d i ,i 1   I ( pi )  I ( pi 1 ) … (6)
i 1 i 1

3.2. Face Detection using Enhanced Fast Fourier Transform for Facial Expression Recognition
In this section the face detection for facial expression recognition is presented. Once after the edges are detected
in the given facial image, local variations in spatial domain is extracted by making use of frequency domain
transform. The unalike brightness, expression dissimilarities and constrictions of face images are denoted in
frequency domain by making use of Enhanced Fast Fourier Transform (E-FFT) that produces the features. The
E-FFT is represented as,
M 1 N 1 2 2
j mk 
 (m, n) e
nl
(k , l )  M N
… (7)
m0 n0

In the equation (7),  ( k , l ) denotes the image co-efficient; ( m, n) denotes the position of RGB value of image
 .Additionally, low frequency components are saved for excerpting discerning features that are required for
face detection. The maximum magnitude value is taken as the feature descriptor of the facial image. This
research work makes use of altered feature representation when comapred with other facee recognition systems
as described in the related work. This work also incorporates a single feature value to represent the entire face
image. Moreover, symbolic items which fit to the same course are denoted by a interval valued variable i.e. by
maximum and minimum magnitude values of N symbolic course of the corresponding face class. For the
symbolic similarity analysis between trained symbolic objects and test object, a new symbolic similarity
measure is devised and employed in this work.

DOI: 10.21817/ijet/2017/v9i6/170906141 Vol 9 No 6 Dec 2017-Jan 2018 4466

ISSN (Print) : 2319-8613
ISSN (Online) : 0975-4024 Dinesh Kunar P et al. / International Journal of Engineering and Technology (IJET)

4. Results and Discussions

Table – 1: Overall Performance Analysis of the proposed IED-FFD with existing works

Elapsed Time for Edge Elapsed Time for Face

Accuracy (in %)
Detection (in seconds) Detection (in seconds) Image
Image Title
LS-KC- Title LS-KC-
GETF [14] IED-FFD IED-FFD IED-FFD
SVM [15] SVM [15]
Image 01 0.563 0.279 0.640 0.405 Image 01 81.63 89.88
Image 02 0.562 0.284 0.678 0.394 Image 02 83.27 90.70
Image 03 0.538 0.238 0.649 0.398 Image 03 81.00 90.45
Image 04 0.556 0.285 0.662 0.423 Image 04 81.99 90.00
Image 05 0.522 0.268 0.640 0.413 Image 05 83.00 90.78
Image 06 0.559 0.236 0.666 0.405 Image 06 82.89 90.50
Image 07 0.512 0.247 0.646 0.438 Image 07 83.55 90.31
Image 08 0.529 0.263 0.669 0.421 Image 08 83.06 89.81
Image 09 0.513 0.288 0.671 0.419 Image 09 83.63 90.90
Image 10 0.517 0.288 0.675 0.444 Image 10 81.21 90.22
Image 11 0.568 0.240 0.657 0.400 Image 11 82.68 91.76
Image 12 0.559 0.288 0.635 0.433 Image 12 82.99 89.13
Image 13 0.532 0.287 0.644 0.433 Image 13 83.18 90.45
Image 14 0.577 0.259 0.685 0.407 Image 14 81.60 89.79
Image 15 0.512 0.278 0.639 0.420 Image 15 82.63 90.24
Image 16 0.541 0.239 0.680 0.385 Image 16 81.70 89.84
Image 17 0.537 0.255 0.662 0.384 Image 17 81.69 89.38
Image 18 0.564 0.285 0.690 0.417 Image 18 81.65 91.33
Image 19 0.566 0.278 0.635 0.435 Image 19 83.65 89.64
Image 20 0.523 0.288 0.657 0.445 Image 20 82.96 89.34
Image 21 0.544 0.269 0.636 0.389 Image 21 81.92 91.06
Image 22 0.541 0.232 0.688 0.420 Image 22 83.80 89.46
Image 23 0.555 0.281 0.630 0.413 Image 23 81.64 91.09
Image 24 0.560 0.286 0.677 0.381 Image 24 81.94 91.52
Image 25 0.563 0.271 0.679 0.404 Image 25 82.08 90.22
Average
Average
Time Taken 0.544 0.268 0.660 0.413 82.45 90.31
Accuracy
(in seconds)
The dataset are obtained from the openly available dataset [16]. 25 images are taken for testing the effectiveness
of the proposed IED-FFD. Performance metrics such as time taken for edge detection, time taken for face
detection and accuracy are taken for comparing the proposed IED-FFD method. In order to compare the time
taken for edge detection, GETF [14] method is chosen for comparison. LS-KC-SVM method is chosen for
comparing the effectiveness of the proposed IED-FFD in terms of time taken for face detection and accuracy.
Table – 1 presents the overall performance analysis of the existing works and proposed work. From the
presented results it is evident that the proposed work IED-FFD consumes less time for edge detection when
compared with GETF [14] method. The average time taken for edge detection of IED-FFD is 0.268 when
compared with GETF [14] method which consumes 0.544 seconds. The results are presented in Fig. 2.
As far as the average time taken for face detection is concerned, the existing LS-KC-SVM [15] had taken 0.660
seconds whereas the proposed IED-FFD consumes 0.413 seconds. The graphical results are presented in Fig. 3.
It is noteworthy that the average accuracy of IED-FFD is 90.31 % which is better when compared with LS-KC-
SVM [15] method that obtains 82.45 %. The results are presented in Fig. 4.

DOI: 10.21817/ijet/2017/v9i6/170906141 Vol 9 No 6 Dec 2017-Jan 2018 4467

ISSN (Print) : 2319-8613
ISSN (Online) : 0975-4024 Dinesh Kunar P et al. / International Journal of Engineering and Technology (IJET)

Original Im
mage Edgge Detected Im
mage Face Detected
D Outpput

Fig.1. (a) Orriginal Image (b) Edge Detected Im

mage (c) Face Deetected Output

Fig. 2. ken for Edge Detection

2 Performane Annalysis: Time Tak

DOI: 10.21817/ijet/2017/v9i6/170906141 Vol 9 No 6 Dec 2017-Jan 2018 4468

ISSN (Print) : 2319-8613
ISSN (Online) : 0975-4024 Dinesh Kunar P et al. / International Journal of Engineering and Technology (IJET)

Fig. 3. Performane Annalysis: Time Tak

ken for Face Deteection

Fig. 4. Performane Analysiss: Accuracy

5. Conclusionn
Facial exxpression recoognition systemm can be deveeloped throug gh schematic process.
p Usuallly the imagess obtained
possibly will have nooises in it. Heence the previious work foccusses on unaaffected seriall prophecy baased filter
techniquee for noise rem moval in faciaal expression recognition im
mages. In thiss second phase of research improved
edge deteection based fast face deteection methodd using enhan nced fourier transform
t (IED
D-FFD) towaards facial
expressioon recongnitioon is proposeed. 25 images are taken fo or evaluating the
t performannce. IED-FFD D has two
folds. Firrst, an improvved edge detecction mechaniism is used that consumes less
l time. Nexxt, fast face deetection is
performeed with improvved accuracy upto 91.76 % with less time.
References
[1] M H. Amer, A. M. Abushaala, "Edge detectionn methods," 2nd World Symposiium on Web Appplications and Networking
G. M.
(WSWWAN), pp. 1-7, 2015.
[2] F. Y. Cui, L. Zou, B. Song,
S "Edge featuure extraction based on digital image processing teechniques," IEEEE International Co
onference on
Autommation and Logisstics, 2008, pp. 23320-2324, 2008.
[3] C. Deeng, W. Ma, Y. Yin,
Y "An edge dettection approach of image fusion based b on improved Sobel operator," 4th Internationnal Congress
on Im
mage and Signal Processing,
P pp. 11189-1193, 2011.
[4] Chinuu, A. Chhabra, "AA hybrid approacch for color basedd image edge detection," Internatiional Conference on Advances in Computing,
Comm munications and Informatics
I (ICA
ACCI), pp. 2443-22448, 2014.
[5] C. X. Deng, G. B. Waang, X. R. Yang, "Image edge deteection algorithm based on improved Canny operatoor," Internationall Conference
on Wavelet
W Analysis and
a Pattern Recoggnition, pp. 168-1172, 2013.
[6] W. Gao,
G X. Zhang, L.
L Yang, H. Liu, "An improved Sobel S edge detecttion," 3rd Internaational Conferencce on Computer Science and
Inform
mation Technologgy, pp. 67-71, 2010.

DOI: 10.21817/ijet/2017/v9i6/170906141 Vol 9 No 6 Dec 2017-Jan 2018 4469

ISSN (Print) : 2319-8613
ISSN (Online) : 0975-4024 Dinesh Kunar P et al. / International Journal of Engineering and Technology (IJET)

[7] B. Du, W. Xiong, J. Wu, L. Zhang, L. Zhang, D.Tao, "Stacked Convolutional Denoising Auto-Encoders for Feature Representation,"
IEEE Transactions on Cybernetics, vol. 47, pp. 1017-1027, 2017.
[8] B.Du, Z.Wang, L.Zhang, L.Zhang, W.Liu, J.Shen, D.Tao, "Exploring Representativeness and Informativeness for Active Learning,"
IEEE Transactions on Cybernetics, vol. 47, pp. 14-26, 2017.
[9] D. Tao, X. Li, X. Wu, S. J. Maybank, "Geometric Mean for Subspace Selection," in IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 31, no. 2, pp. 260-274, Feb. 2009.
[10] D. Tao, X. Li, X. Wu, S. J. Maybank, "General Tensor Discriminant Analysis and Gabor Features for Gait Recognition," IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 29, pp. 1700-1715, 2007.
[11] C. Peng, J. Cheng, Q. Cheng, “A supervised learning model for high-dimensional and large-scale data,” ACM Transactions on
Intelligent Systems and Technology (TIST), vol. 8, pp. 1 – 23, 2017.
[12] Q. Cheng, H. Zhou, J. Cheng, H. Li, "A Minimax Framework for Classification with Applications to Images and High Dimensional
Data," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 36, pp. 2117-2130, 2014.
[13] P.S. Hiremath, C.J. Prabhakar, “Face Recognition Using Symbolic KPCA Plus Symbolic LDA in the Framework of Symbolic Data
Analysis: Symbolic Kernel Fisher Discriminant Method,” International Conference on Advanced Concepts for Intelligent Vision
Systems, pp. 982–993, 2008.
[14] H. Roy, D. Bhattacharjee, “Heterogeneous face matching using geometric edge-texture feature (GETF) and multiple fuzzy-classifier
system,” Applied Soft Computing, vol. 46, pp. 967 – 979, 2016.
[15] Q. Q. Tao, S. Zhan, X. H. Li, T. Kurihara, “Robust face detection using local CNN and SVM based on kernel combination,”
Neurocomputing, vol. 211, pp. 98-105, 2016.
[16] image dataset: https://github.com/StephenMilborrow/muct
[17] P. Dinesh Kumar, Dr. B. Rosiline Jeetha, “Unaffected Serial Prophecy based Filter Technique for Noise Removal in Facial Expression
Recognition Images,” International Journal of Civil Engineering and Technology (IJCIET), vol. 8, pp. 497 – 506, 2017.
[18] P. Ekman, T. Huang, T. Sejnowski, J. Hager, "Final Report to NSF of the Planning Workshop on Facial Expression Understanding",
1992.
[19] EwaPiatkowska "Facial Expression Recognition System", Master's Thesis Technical Report.

DOI: 10.21817/ijet/2017/v9i6/170906141 Vol 9 No 6 Dec 2017-Jan 2018 4470