View
2
Download
0
Category
Preview:
Citation preview
Gender Recognition Using Facial Images
1 Department of Computer Engineering, Fırat University Elazig, Turkey
2 Department of Computer Education and Instructional Technology, Siirt University Siirt, Turkey
Abstract. In this study, gender classification is performed based on front façade photos of 100 male and
100 female. In order to demonstrate the internal face images are aligned and cropped.. Even though some
images are cropped about ears and hairs with the expense of the information loss about gender information at
those parts, the main aim is achieving gender classification on internal face of the human body. It has been
generated that 7 x 200 matrix which obtained from images that include 3 statistical values (average, standard
deviation and entropy) and 4 parameters of GLCM (Gray Level Co-occurrence Matrix). 60% gender
classification accuracy rate is achieved based on the generated frontal face image data set . As a secondary
method, features are extracted by means of GLCM method, followed by application of 2D DWT (The
Discrete wavelet transform) technique on the original images. it has been established attribute of original
images by respectively DWT (The Discrete wavelet transform) and GLCM (Gray Level Co-occurrence
Matrix). When first method is used for 7 photos (7 attitude) which are output 2D DWT, set volume is 49 x
200. Used to 5 different wavelets of relatives and the highest achievement is found at Coiflets Wavelets Filter
by 88%. Second method increases to first method's achievement by %46.
Keywords: Gray Level Co-occurrence Matrix, Discrete wavelet transform, gender classify, facial images
1. Introduction
Gender recognition is the highly effortless cognition among people but it is very complicated process for
the computer. For social life, gender factor undertakes effective role in the communication. Computer based
system in which automatic gender recognition process is a field of the computer vision. This process is
executed with the facial informs or any parts of body which exclude such informs. At the process of facial
informs, due to the fact that peculiar feature of gender such as make-up or beard decrease to similarity ratio,
facilitate to identify and increase reliability and robust of system classify. It has been studied that a lot of
survey for gender recognition with computer system[1]-[5]. Feature extraction and step of classify are
investigated at such studies. Thanks to find knowledge about less but distinctive property, Feature Extraction
Method aids to classify. Basically, classify process is a separated form of a group data where act to similar
mission. Generally, we find two types of gender classify in related studies. One of them is Global Feature-
based and the other is Geometric Feature-based [1]. While Global Feature-based study on the training images,
Geometric Feature-based arrange with about body parts such as nose, ears and hairs[6]. It has been studied
that a lot of survey about method of identifying the gender feature and classifying such methods. These
surveys were started by Golomb [7], Cottrel and Metcalfe [8] at 1990s. In both studies, images are aligned by
hand and later run with the Multi-Layer Perceptron (MLP) directly. Thus, classify process is actualized. Not
to be used Feature Extraction Methods, all pixels of images are adopted as feature and it has been obtained
that classify achievement by %92. Mozaffari et al. [9] have combined global and local feature in their study.
It has been used Discrete Wavelet Transform (DWT) and Local Binary Pattern (LBP) for extract to Global
Feature-based that obtained classify achievement by %85. And concluded that female faces are extended and
Corresponding author. Tel.: + (90) 555 417 93 76 ; fax: +(90) 484 223 19 98
E-mail address: serdarabut@gmail.com
2013 International Conference on Agriculture and Biotechnology
IPCBEE vol.60 (2013) © (2013) IACSIT Press, Singapore
DOI: 10.7763/IPCBEE. 2013. V60. 22
112
Burhan ERGEN1
and Serdar ABUT2
circular than male faces. Han [10] used to 3D GavabDB for extracts to Geometric Feature-based which is
belonging to face.
It has been determined that male and female’s basic and distinctive feature of faces. Man’s eyebrow is
straight and thick when it compares to female and female have smaller nose than male. GavabDB data set
which consists of 427 images and SVM Classification are used, thus, it has been obtained classify
achievement by %82.56.
Rahman Khorsandi [11], studied on the 2D ears image that is the first study in the gender recognition
fields. He used to Gabor filter for extract feature. In a study, Notre Dame University has obtained classify
achievement by %89.49 in which is used to J Data Group.
In this study, aligned and cropped gender identify is processed. And creation of face database
information will be given at next part. It has been mentioned that feature group in which obtained original
photos in GLCM parameter at part 3. Extract feature is obtained from DWT outcomes as new images and
considered at part 4. Analyse detail of method and compare of each other will be given part 5. And finally it
will be discussed study results and planning future ideas.
2. Creation Of Image Data Set
In this study, against the front of the photos (cropped and aligned 100 male, 100 female) has been used
which are obtained from FEI Face Database. Fig. 1. shows that example FEI faces in the database. While top
row of Fig. 1 shows the male face, the bottom row composes to the female face.
Fig. 1: Some faces in FEI database.
Feature group has been extracted thanks to apply on GLCM from original condition in photos. It has
been extracted that a new feature set which is obtained from new images. Consisting of 2D wavelet
transform with GLCM have been used in this process. When creating of this feature set, 103 filters in which
a relative with 5 different wavelet filter have been extracted for individual feature sets. Providing the highest
performance of the wavelet filter has been searched by iterative methods. For created the two feature set, it
has been used SVM classifier (10 fold CV) and classification performance of these methods have been
compared.
3. Pre-processing
3.1. Discrete Wavelet Transform
DWT has been used in many signal and image processing applications [12]-[15]. An image in which 2D
wavelet transform is expressed to a wavelet function and a scaling function in terms of conversions and
evaluations. This functions are calculated by 2D filter groups where contain a low and a high pass. After 2D
resolution, the image as shown in Fig. 2 is divided to components of multi-resolution in to individual
frequencies.
Sub-bands are referred to as details. k; the scale and J; the largest scale in
the separate. While the original image size is M x N, sub-band size which owns k scale ⁄⁄ . Sub-
band , the lowest resolution residue and j; is selected range from ⁄⁄ and
[16]. Fig. 2 shows sub-bands for a sample photo which are obtained after the 2D Wavelet
Transform. When 2D DWT is applied to an image in data set (Fig. 3), 7 photos is obtained.
113
Fig. 2: Sub-bands of 2D
orthogonal Wavelets Transf. Fig. 3: Sub-images are obtained
from a facial image in dtbs
4. Future Extraction
After pre-processing, it has been tried to extract the Gray Level Co-occurrence Matrix of images and
local binary pattern feature.
4.1. Gray Level Co-occurrence Matrix
GLCM is one of the most well known method of texture analysis and it would predict image feature
related to second-order statistics [17]. GLCM is defined as follows: Point (i,j) in GLCM represents frequency
of such pixels that is within a certain window, in the direction of , d from adjacent distance. For d, 1 or 2
is selected. takes four values to and . As an example is shown Fig. 4; pixel values are 1
and 2, for ; frequency value is 3, Gray Level Co-occurrence Matrix of 1 and 2 marks field
value is 3. GLCM's each element is calculated as follows;
∑ ∑
Where; ; frequency , ; direction, d; distance between two adjacent pixels (pixel values are
and )
Fig. 4: An example of GLCM is obtained from an original image.
GLCM reflects to the gray information of images by synthesizing such as level of color and changing in
the direction and distribution [18]. Derived parameters which are obtained by GLCM, can be used for feature
extraction. Haralick et al. have found that 14 varieties form distinguishing feature about based on GLCM
image texture feature which consist of quantitative GLCM description method [19] and [20]. It has been used
4 parameters of GLCM in this study. These parameters include;
Average, standard deviation and entropy values are also added to GLCM feature and 7 features have
been obtained totally. In the study, used 200 photos which for each of them, it has been used above 7 features
and created a data set which is named GLCM. Then, 2D DWT is implemented and it has been created that 7
photos as shown at Fig. 4. It has been extracted 7 GLCM features for each of these 7 photos. So, a total of
7x7 = 49 values for each photo have been obtained. Thus, about 200 photos, it has been created another data
set that consists of 200 x 49 = 9800 features and is named DWT_GLCM.
114
Energy: ∑
Homogeneity: ∑ | | ⁄
Contrast : ∑ | |
Correlation: ∑
5. Experimental Results
It has been specified that 4 images of GLCM parameter in first method which mentioned at Chapter 4.1.
In addition, the standard deviation, the entropy and the average values have been calculated. Thus, relating to
200 photos that consist of 100 male and 100 female photos, 7 features have been extracted for each image
and 200 x 7 feature set is created. Then, it is named GSEM feature set, when it is classified as gender and
support from 10-fold SVM (Support Vector Machines), achievement is by %59.5.
It has been tried iterative study about different relative filters of 2D DWT in the second method. Then it
has been classified with the feature set separately. This classification achievements Fig. 5-Fig. 9 are shown
separately.
Fig. 5: Achievements of classify data set which are
created by 45 Daubechies wavelet filters (db1 – db45)
Fig. 7: Achievements of classify data set which are
created by 23 Symlets wavelet filters (sym2 – sym24)
Fig. 9: Achievements of classify data set which are
created by 15 Reverse Biorthogonal wavelet filters
Fig. 6: Achievements of classify data set which are created
by 5 Coiflets wavelet filters (coif1-coif5)
Fig. 8: Achievements of classify data set which are created
by 15 Biorthogonal wavelet filters
Fig. 10: Averages achievement of classify which are created
by 5 wavelet filter families
Averages of classification results show Fig. 10 which are taken by classification of feature sets. Feature
sets are obtained from 103 kinds filter in which 5 wavelet families. Following the implementation of the
various filters to images, coif1 filter is highest result which has been obtained with family Coiflet filter (Fig.
6). 1 is the reset moment value in that scaling and wavelets functions. Thanks to Coif1 filter, gender
classification performance is increased up to 89%. The average is 83% where wavelet transform in all of
these trials.
6. Results and Discussion
0.76
0.78
0.80
0.82
0.84
0.86
0.88
0 20 40
Daubechies wavelets Correct Rate
0.80
0.82
0.84
0.86
0.88
0 5 10 15 20 25
Symlets Correct Rate
0.80
0.82
0.84
0.86
0.88
0 5 10 15
Reverse biorthogonal wavelets Correct Rate
0.80
0.82
0.84
0.86
0.88
0.90
0 2 4 6
Coiflets Correct Rate
0.78
0.80
0.82
0.84
0.86
0 5 10 15
Biorthogonal wavelets Correct Rate
db coif sym bior rbio General 0.79
0.80
0.81
0.82
0.83
0.84
0.85
1
115
In these studies, it is investigated that the impact on gender recognition of 2D wavelet transform. GLCM
and 3 statistical values have been applied on the original image and 7 features are studied in all cases. While
200x7 feature set is used in the first method, 200x49 feature set is used in the second method due to 7 photos
in which the outputs of the wavelet transform. While it has been achieved that in 60% with 7 feature in the
first method, the second method has been achieved in 89% with 49 feature. In this way there is provided a 48%
increase in success.
In future studies, thanks to feature selection methods, these 49 DWT features which are created with
different filters will be reduced to 7. In this way, gender characteristics on which applying to wavelet
transform of the image will be measured exactly. Basically, there are two reasons for increase achievement
here. One of them is the expansion of the feature set, the other is the effect of DWT.
At the feature extraction and the classification processes, due to the fact that best performance is aimed
to with at least features, same performance will be struggled in achieve with reducing the number of feature.
7. References
[1] Cellerino, A.,Borghetti, D., &Sartucci, F., “Sex differences in face gender recognition in humans”, Brain Research
Bulletin, 63, 443-449, 2004.
[2] Ullah, I.;Hussain, M.; Muhammad, G.; Aboalsamh, H.; Bebis, G.; Mirza, A.M., "Gender recognition from face
images with local WLD descriptor,"Systems, Signalsand Image Processing (IWSSIP), 2012 19th International
Conference on , vol., no., pp.417,420, 11-13 April 2012
[3] Tizita N. S.,“Age group and gender recognition from human facial images”, Ethiopian Society of Electrical
Engineers 6th Scientific Conference on Electrical Engineering (CEE-2012), 2012
[4] Kekre H.B., Sudeep D. T., Tejas C.“Face and Gender Recognition Using Principal Component Analysis”,(IJCSE)
International Journal on Computer Science and Engineering Vol. 02, No. 04, 959-964, 2010.
[5] Yasmina A., Ramón A. M., Pedro G.S.,“Gender Recognition from a Partial View of the Face Using Local Feature
Vectors”, Pattern Recognition and Image Analysis Lecture Notes in Computer Science Volume 5524, pp481-488,
2009.
[6] Khan, S.A.; Nazir, M.; Akram, S.; Riaz, N., "Gender classification using image processing techniques: A survey,"
Multitopic Conference (INMIC), 2011 IEEE 14th International , vol., no., pp.25,30, 22-24 Dec. 2011.
[7] B. Golomb, D. Lawrence, and T. Sejnowski, "Sexnet: A Neural Network identifies Sex from human
faces,"Advance in neural information processing systems, pp. 572-577, vol. 3, 1990.
[8] G. Cottrell and J. MetcaJfe,"Empath: Face, emotions and gender recognition using Holons,"Neural information
processing systems, pp. 564-571.vol. 3, 1990.
[9] S. Muzaffari, H. Behravan and R. Akbari, "Gender classification using single frontal image per person,"IEEE
International Conference on Pattern Recognition, pp. 1192 - 1195, August. 2010.
[10] X. Han, H. Ugail and I. Pahnar, "Gender classification based on 3D face geometry features using SVM",IEEE
international conference on cyber world, page. 114 -118, September 2009.
[11] Khorsandi, R.; Abdel-Mottaleb, M., "Gender classification using 2-D ear images and sparse representation,
"Applications of Computer Vision (WACV), 2013 IEEE Workshop on , vol., no., pp.461,466, 15-17 Jan. 2013.
[12] Ricardo J. C., Rafael G., Francisco J. B., Marcos M., “Flexible architecture for the implementation of the two-
dimensional discrete wavelet transform (2D-DWT) oriented to FPGA devices”, Microprocessors and
Microsystems, Volume 28, Issue 9, Pages 509-518, 2 November 2004.
[13] Sourour K., Ridha D., Rached T., “Efficient hardware architecture of 2D-scan-based wavelet watermarking for
image and video”, Computer Standards & Interfaces, Volume 31, Issue 4, Pages 801-811, June 2009.
[14] Rinky B.P., Payal M., Manikantan K., Ramachandran S., “DWT based Feature Extraction using Edge Tracked
Scale Normalization for Enhanced Face Recognition”, Procedia Technology, Volume 6, Pages 344-353, 2012.
[15] Maged M. M. F., “Online handwritten signature verification system based on DWT features extraction and neural
network classification”, Ain Shams Engineering Journal, Volume 1, Issue 1, Pages 59-70, September 2010.
[16] Haifeng H., “Variable lighting face recognition using discrete wavelet transform”, Pattern Recognition Letters,
116
Volume 32, Issue 13, Pages 1526-1534, 1 October 2011.
[17] Zhu L., Zhang Z., "Auto-classification of insect images based on color histogram and GLCM," Fuzzy Systems and
Knowledge Discovery (FSKD), 2010 Seventh International Conference on , vol.6, no., pp.2589,2593, 10-12 Aug.
2010.
[18] Jia L., Zhou Z., Li B., "Study of SAR Image Texture Feature Extraction Based on GLCM in Guizhou Karst
Mountainous Region," Remote Sensing, Environment and Transportation Engineering (RSETE), 2012 2nd
International Conference on, vol., no., pp.1,4, 1-3 June 2012.
[19] Haralick R.M., Shanmugam K., Dinstein I., "Textural Features for Image Classification," Systems, Man and
Cybernetics, IEEE Transactions on, vol.SMC-3, no.6, pp.610,621, Nov. 1973.
[20] Anys H., Bannari A., He D. C., Morin D., “Texture analysis for the mapping of urban areas using airborne MEIS-
II images”, Proceedings of the First International Airborne Remote Sensing Conference and Exhibition, Pages
231-245, 1994.
117
Recommended