You are on page 1of 7

IJCSN International Journal of Computer Science and Network, Volume 5, Issue 5, October 2016

ISSN (Online) : 2277-5420 www.IJCSN.org


Impact Factor: 1.02
763

Face Recognition Using SOM Neural Network


with DCT-PT Facial Feature Extraction Techniques
1
Mahendra Kumar, 2 Rohit Giri , 3
Shipla Jangid

Abstract - The paper deals with 5 different techniques for recognition system. As a result, to improve the
feature extraction of face. First-step in face recognition recognition performance of such systems one has to
systems is face detection, with purpose of localizing and enhance these operations. Prior to classification, the
extracting the face region from the background. Self- features should be extracted from the human face images.
Organizing Map (SOM) Neural Network has been used for Feature extraction in the sense of some linear or nonlinear
training of database and simulation of FR system. The
transforms of the face images with subsequent feature
algorithm developed for the face recognition system and an
image-based approach is formulated, using Directional selection is commonly used for reducing the
Discrete Cosine Transform (DDCT), Discrete Wavelet dimensionality of these images so that the extracted
Transform (DWT), Discrete Cosine Transform (DCT) , features are as representative as possible. However, the
Sobel Edge Detection and DCT-Pyramid Transform (DCT- problem of extracting features from a human face remains
PT) , simulated in MATLAB. Simulation results are very a barrier to apply the practical applications, since it is
promising. influenced by the lighting condition, illumination
changes, various backgrounds and individual variations.
Keywords - Face Recognition (FR), Directional Discrete
Cosine Transform (DDCT), Discrete Cosine Transform (DCT), Recently, several methods for feature extraction have been
Discrete Wavelet Transform (DWT), Sobel Edge Detection proposed [1-5]. Of these reported techniques, the principle
(SED), SOM Neural Network, DCT-PT.
component analysis (PCA)) [1] and linear discriminant
analysis (LDA) [2] are regarded to be the well known
1. Introduction feature extraction methods. PCA linearly projects the
image space along the direction of maximal variance into
The automatic recognition of people is a challenging a low-dimensional subspace. These methods are statistical
problem which has received much attention during the linear projection methods which largely rely on the
recent years due to its many applications in different representation of the training samples. Major drawbacks
fields. Face recognition is one of those challenging with these techniques are the performance drop of face
problems and up to date, there is no technique that recognition whenever face appearances are subject to
provides a robust solution to all situations and different variations by factors such as illumination, expression,
applications that face recognition may encounter. Face pose, accessories and aging. Moreover, they require
recognition has several characteristics that are intensive computations in determining the basis space for
advantageous for consumer applications. In addition to, a large number of images and when adding new images to
the need for an automatic face recognition system a facial image database. On the other hand, multi-
especially in the field like at the border control, airports resolution techniques have been used for feature
etc. is becoming very important to strengthen the security extraction in many face recognition systems [58].
check. Generally, feature extraction and classification Among them, the most popular are the discrete wavelet
criterion are the two basic operations of any face transform (DWT). 2D DWT has been used in various face

Rohit Giri is M.Tech Scholar with the Department of Electronics & Communication Engineering, Mewar University, Chittorgrah , Rajasthan ,India (e-
mail: goswami.rohit41@gmail.com).
Mahendra Kumar is Faculty with Department of Electronics & Communication Engineering, University College of Engineering, RTU, Kota, India.
(Corresponding author e-mail: miresearchlab@gmail.com).
Shilpa Jangid is Faculty with the Department of Electronics & Communication Engineering, Mewar University, Chittorgrah , Rajasthan ,India ( e-mail
:shilpa.jangid08@gmail.com).
IJCSN International Journal of Computer Science and Network, Volume 5, Issue 5, October 2016
ISSN (Online) : 2277-5420 www.IJCSN.org
Impact Factor: 1.02
764

recognition systems in order to extract multiple sub-band


face images. These sub-band images contain coarse
approximations of the face as well as horizontal, vertical
and diagonal details of faces at various scales. These
wavelet-based methods focus on the sub-bands that
contain the most relevant information to better represent
the face image.

Further, the discrete cosine transform (DCT) has been


employed in face recognition [9][12]. The DCT has
several advantages over the PCA, two of them are the first
one is, DCT is data independent and the second one is
DCT can be implemented using a fast algorithm. The
discrete cosine transform (DCT) represents an image as a
sum of sinusoids of varying magnitudes and frequencies.
The DCT has the property that, for a typical image, most
of the visually significant information about the image is
concentrated in just a few coefficients of the DCT. For
this reason, the DCT is often used in image compression Fig 1: Block diagram of the face recognition system [15]
applications. For example, the DCT is at the heart of the
international standard lossy image compression 2.1 Pre-processing
algorithm known as JPEG.
The first step in any face recognition system is Face
The other one technique is Sobel Edge detection is the Detection. Many face detection techniques have been
process of localizing pixel intensity transitions[20-30]. proposed in the past decade. They can be classified into
The edge detection have been used by object recognition, geometry-based face detectors and colour-based face
target tracking, segmentation, and etc. Therefore, the detectors. Among the geometry-based face detectors, a
edge detection is one of the most important parts of image method examines the triangle-relationship between eye
processing. The remainder of the paper is organized as and mouth regions to identify the face region. In addition
follows: In Section II, The proposed face recognition to, the traditional eye detection methods can be simply
method is briefly explained. The SOM neural network and efficiently implemented for frontal face images but
system is described in Section III. Experimental results can be difficult for complex images. Moreover, skin
and conclusions are presented in Sections IV and V, colour has been proven to be an effective image feature for
respectively. face detection. The automatic face detectors based on skin
colour are relatively fast and therefore are beneficial for
2. Face Recognition Method consumer applications such as real-time face recognition
embedded in a smart home environment. In the pre-
Block diagram of the proposed face recognition system is processing stage of the proposed system, a facial region
shown in fig. 1. In this figure the three key steps of face based on skin colour detection is cropped from an input
recognition, namely, face detection, facial feature image. The obtained facial region is then resized into an
extraction and face recognition are identified. Detailed 88 pixel image to make the face recognition system scale
descriptions of these steps are addressed in the following invariant. Then after, histogram equalization is applied to
sub-sections[15-22]. enhance the image brightness and contrast.

2.2 Facial Feature Vector Extraction

For searching results there are so many technique


available for facial extraction .We are dealing with feature
extraction methods which are based on the transforms
such as the DWT, DCT and Sobel edge detection also
commonly represent the face images with a large set of
features. The features of every image stored in our data
based are extracted and then stored into the feature vector.
IJCSN International Journal of Computer Science and Network, Volume 5, Issue 5, October 2016
ISSN (Online) : 2277-5420 www.IJCSN.org
Impact Factor: 1.02
765

Once the feature vectors for all existing images are n1 = IW11 p
developed the new database consist of all feature vectors
is formed and then stored inside our storage device. To a1 = compet(n1 )
retrieve all images that are similar to the target image, we
must extract the features of the target image and compare
it with all features vectors. Discrete cosine transform So, when a vector p is presented, the weights of the
(DCT) is a powerful transform to extract proper features winning neuron and its close neighbours move toward p.
for face recognition. After applying DCT to the entire face Consequently, after many presentations, neighbouring
images, some of the coefficients are selected to construct neurons learn vectors similar to each other[16]. Hence,
feature vectors. the SOM network learns to categorize the input vectors it
sees.
3. Self-Organizing Map (SOM) Neural
Network The SOM network used here contains N nodes ordered in
a two-dimensional lattice structure. In these cases, each
3.1 Network Architecture node has 2 or 4 neighboring nodes, respectively.
Typically, a SOM has a life cycle of three phases: the
In Network Architecture , SOMs can be one-dimensional, learning phase, the training phase and the testing phase.
two-dimensional or multi-dimensional maps. The number
of input connections in a SOM network depends on the IV. Proposed DCT Based Laplacian Pyramid
number of attributes to be used in the classification [14]. Transform
Laplacian pyramid: The Laplacian pyramid was first
introduced as a model for binocular fusion in human
stereo vision [3], where the implementation used a
Laplacian pyramid and a maximum selection rule at each
point of the pyramid transform. Essentially, the procedure
involves a set of band-pass copies of an image is referred
to as the Laplacian pyramid due to its similarity to a
Laplacian operator. Each level of the Laplacian pyramid
is recursively constructed from its lower level by applying
the following four basic steps: blurring (low-pass
filtering);sub-sampling (reduce size); interpolation
Fig 2: Architecture of a simple SOM Neural Network [15] (expand); and differencing (to subtract two images pixel
by pixel). In the Laplacian pyramid, the lowest level of
The input vector p shown in fig. 2 is the row of pixels of the pyramid is constructed from the original image [5].
the input compressed image. The ||dist|| box accepts the
input vector p and the input weight matrix IW1, 1, which Discrete Cosine Transform (DCT)
produces a vector having S1 elements. The elements are
the negative of the distances between the input vector and The Discrete cosine transform (DCT) is an important
vectors iIW1,1 formed from the rows of the input weight transform in image processing. Large DCT coefficients
matrix. The ||dist|| box computes the net input n1 of a are concentrated in the low frequency region; hence, it is
competitive layer by finding the Euclidean distance known to have excellent energy compactness properties.
between input vector p and the weight vectors. The The 2D discrete cosine transform Z (u, v) of an image or
competitive transfer function C accepts a net input vector 2D signal z(x, y) of size MxN is define as:
for a layer and returns neuron outputs of 0 for all neurons
except for the winner, the neuron associated with the most
positive element of net input n1. The winners output is 1.
The neuron whose weight vector is closest to the input
vector has the least negative net input and, therefore, wins
the competition to output a1. Thus the competitive transfer
function C produces a 1 for output element a1i
corresponding to i*, the winner. All other output
elements in a1 are 0[16].
IJCSN International Journal of Computer Science and Network, Volume 5, Issue 5, October 2016
ISSN (Online) : 2277-5420 www.IJCSN.org
Impact Factor: 1.02
766

Directional Discrete Cosine Transforms (D-DCT): In


this method, first, we have computed 2D-DCT of a word
image. Let he 2D-DCT coefficient matrix of a word image
be D. Then, principal diagonal, upper N 2 and lower N
2 diagonals of D is extracted before and after flipping
D, and further computed their standard deviations
respectively as discussed in the aforementioned
paragraph. Meanwhile, we have also extracted features
based on conventional DCT. In this case, its coefficient
matrix is divided into four zones and standard deviation
of each zone is computed. Thus, a feature vector of four
features is formed[26].

4. Experimental Results
4.1 Image Database

Face image database was created for the purpose of


benchmarking the face recognition system. The image
DCT based Laplacian Pyramid transform: The database is divided into two subsets, for separate training
procedure for Laplacian pyramid construction and and testing purposes. During SOM training, 30 images
reconstruction is illustrated in Fig-I. were used, containing six subjects and each subject having
5 images with different facial expressions. fig 3 shows the
Reduced Function: Image at the Oth level go of size training and testing image database constructed.
MxN is reduced to obtain next level g, of size 0.5MxO.5N
where both spatial density and resolution are reduced.
Similarly, g2 is the reduced version of g, and so on. Image
reduction is done by taking the DCT and applying the
mCT on first half of coefficients in both directions. The
level to level image reduction is performed using the
function reduce[23-30].

Fig 3: Image database for training [15]

Face recognition system presented in this paper was


developed, trained, and tested by using MATLAB 7.5.
The computer was a Windows 8 with a 2.40GHz Intel(R)
core(TM) i3-3110M processor and 2 GB of RAM.

4.2 Validation of Technique

Pre-processed grayscale images of size 8 8 pixels are


reshaped in MATLAB to form a 64 1 array with 64
Fig. DCT based Laplacian Pyramid Transform rows and one column for each image. This technique is
IJCSN International Journal of Computer Science and Network, Volume 5, Issue 5, October 2016
ISSN (Online) : 2277-5420 www.IJCSN.org
Impact Factor: 1.02
767

performed on all 5 test images to form the input data for TABLE 1
FACE RECOGNITION RATE AT EPOCH=1000
testing the recognition system. Similarly, the image
Techniq Test Test Test Test Test Test Recogn
database for training uses 30 images and forms a matrix ue Img 1 Img 2 Img 3 Img Img Img 6 ition-
of 64 30 with 64 rows and 30 columns. The input 4 5 rate
vectors defined for the SOM are distributed over a 2D- SED NR R R R R R 84.33%
input space varying over [0 255], which represents
DCT R R R R R R 100%
intensity levels of the gray scale pixels. These are used to
train the SOM with dimensions [64 2], where 64 DWT R R R R R R 100%
minimum and 64 maximum values of the pixel intensities DDCT R R R R R R 100%
are represented for each image sample. The resulting DCT- R R R R R R 100%
SOM created with these parameters is a single-layer feed PT
forward SOM map with 128 weights and a competitive
transfer function. The weight function of this network is 5. Conclusion
the negative of the Euclidean distance [13]. As many as 5
test images are used with the image database for This paper presents a novel face recognition technique
performing the experiments. Training and testing sets that uses features derived from DCT-PT, DDCT, DCT,
were used without any overlapping Fig. 4 shows the result DWT, Sobel coefficients, along with a SOM-based
of training and testing simulated in MATLAB using the classifier. The system was evaluated in MATLAB using
image database and test input image. an image database of 30 face images, containing six
subjects and each subject having 5 images with different
facial expressions. After training for approximately 1000
epochs the system achieved a recognition rate of as shown
in table 1 for 5 consecutive trials. A reduced feature
space, described for experiment, dramatically reduces the
computational requirements of the methods. DCT-PT
feature extraction method gives better results compared to
DDCT, DWT, DCT and SED methods as shown in table
2. This makes our system well suited for high speed, low-
cost, real-time hardware implementation.

Fig 5: For DDCT: (a) SOM Layer Vector (b) Weight Vector
IJCSN International Journal of Computer Science and Network, Volume 5, Issue 5, October 2016
ISSN (Online) : 2277-5420 www.IJCSN.org
Impact Factor: 1.02
768

Analysis and Machine Intelligence, vol. 22, no. 6, pp.


570-582, 2000.
[12] K. H. Tan and M. Ghanbari Layered image coding
using the DCT pyramid, IEEE Trans. on Image
Processing, vol. 4, no. 4, pp. 512-516, April 1995.
[13] AYBAR, E., Topolojik Kenar _slecleri, Anadolu
(d) niversitesi, Fen Bilimleri Enstits, Ph.D. thesis,
2003.
[14] Image Toolbox (for use with Matlab) Users Guide,
The MathWorks Inc., 2000.
[15] J. Nagi, Design of an Efficient High-speed Face
Recognition System,Department of Electrical and
Electronics Engineering, College of Engineering,
Fig 6: Best match image of subject found in training database (a) Face Universiti Tenaga Nasional, March 2007.
recognition using DWT (b) Face recognition using DCT (c) Face
recognition using SOBEL edge detection (d) Face recognition using DDCT [16] A. Abdallah, M. Abou El-Nasr, and A. Lynn Abbott,
(e) Face recognition using DCT-PT. A New Face Detection Technique using 2D DCT
and Self Organizing Feature Map in Proc. of World
Academy of Science, Engineering and
References Technology,Vol. 21, May 2007, pp. 15-19.
[17] Y. Zi Lu and Z. You Wei, Facial Expression
[1] M. A. Turk and A. P. Pentland, Face recognition Recognition Based on Wavelet Transform and MLP
using eigenfaces, Proc. of IEEE Conference on Neural Network, in Proc. 7th International
Computer Vision and Pattern Recognition, pp. 586- Conference on Signal Processing, ICSP 2004, Vol. 2,
591, June 1991. Aug 2004, pp. 1340-1343.
[2] X. He, S. Yan, Y. Hu, P. Niyogi, and H. Zhang, Face [18] Nisha Soni, Garima Mathur and Mahendra Kumar,
recognition using Laplacian faces, IEEE Transactions Face Recognition using SOM Neural Network with
on Pattern Analysis and Machine Intelligence, vol. 27, Different Facial Feature Extraction Techniques
no. 3, pp. 328-340, 2005. International Journal of Computer Applications
[3] Y. Bai, L. Lianwen Jin, and Q. Huang, Novel face (IJCA), Vol. 76, No. 3, Aug. 2013, pp.7-11. (Impact
recognition approach based on steerable pyramid factor: 0.821)
feature, IEEE International Conference on ICIP2009, [19] Nisha Soni, Garima Mathur and Mahendra Kumar,
Cairo, Egypt, pp. 4165-4168, 2009. A Matlab Based High Speed Face Recognition System
[4] Z. M. Hafed and M. D. Levine, Face recognition Using SOM Neural Networks International Journal
using the discrete cosine transform, International
of Engineering Research and Applications (IJERA),
Journal of Computer Vision, vol. 43, no. 3, pp. 167-
Vol. 3, Issue 4,July-Aug. 2013, pp.785-790. (Impact
188, 2001.
factor: 1.69)
[5] H. K. Ekenel and B. Sankur, Mult-resolution face
[20] M. Kumar et.al., Digital Image Watermarking using
recognition, Image and Vision Computing, vol. 23,
Fractional Fourier transform via image compression,
pp. 469477, 2005.
In IEEE International Conference on Computational
[6] C. Garcia, G. Zikos, and G. Tziritas, Wavelet packet
Intelligence and Computing Research 2013 (IEEE
analysis for face recognition, Image and Vision
ICCIC-2013), 26-28 Dec., 2013.
Computing, vol. 18, no. 4, pp. 289297, 2000.
[21] Mahendra Kumar et. Al..Fuzzy type Image Fusion
[7] L. Shen and L. Bai, A review on Gabor wavelets for
using SPIHT Image compression technique,
face recognition, Pattern Analysis and Applications,
International Journal of Engineering Research and
vol. 9, no. 2, pp. 273-292, 2006.
Applications (IJERA), (Accepted). (Impact factor:
[8] D. Kim, I. Jeon, S. Y. Lee, P.K. Rhee, and D. J.
1.69)
Chung, Embedded face recognition based on fast
[22] Mahendra Kumar et. al., Digital Image
genetic algorithm for intelligent digital photography,
Watermarking using Fractional Fourier Transform
IEEE Transactions on Consumer Electronics, vol. 52,
with Different Attacks International Journal of
no. 3, August 2006.
Scientific Engineering and Technology, Volume No.3
[9] D. Koller, and M. Sahami, Towards optimal feature
Issue No.8, Aug. 2014, pp : 1008-1011. ((ISSN : 2277-
selection, In ICML1996, Bari, Italy, pp. 8795, 1996.
1581))
[10] M. Raymer, W. Punch, E. Goodman, L. Kuhn, and A.
[23] Rajesh Kumar Kakerda et. Al.,Fuzzy type Image
Jain, Dimensionality reduction using genetic
Fusion using hybrid DCT-FFT based Laplacian
algorithms, IEEE Transactions on Evolutionary
Pyramid Transform, 4th IEEE International
Computation, vol. 4, no 2, pp. 164-171, 2000.
Conference on Communication and Signal Processing
[11] C. Liu and H. Wechsler, Evolutionary pursuit and its
(ICCSP 2015) 02-04 April 2015 - Melmaruvathur,
application to face recognition, IEEE Trans. Pattern
TN, IND.
IJCSN International Journal of Computer Science and Network, Volume 5, Issue 5, October 2016
ISSN (Online) : 2277-5420 www.IJCSN.org
Impact Factor: 1.02
769

[24] M. Kumar et. Al., Comparative Study Of Different [28] Mahendra Kumar et. al., Image Fusion based on the
Classifiers Based Speaker Recognition System Using Modified Curvelet Transform In International
Modified MFCC For Noisy Environment, Conference on Smart Trends for Information
International Conference Green Computing and Technology and Computer Communications
Internet of Things (ICGCIoT - 2015) 08-10 Oct., 2015, (SmartCom 2016), Jaipur, India, 6 7 August, 2016.
Delhi, IND. (Proceeding in CCIS Series of Springer ) (ISSN
[25] M. Kumar et. Al., Robust Digital Image Number - 1865-0929).
Watermarking using DCT based Pyramid Transform [29] Mahendra Kumar et. Al., Robust Image Fusion
via image compression, 4th IEEE International based on Optimal Cuve-let Transform, IEEE
Conference on Communication and Signal Processing International Conference on Micro-Electronics and
(ICCSP 2015) 02-04 April 2015 - Melmaruvathur, Telecommunication Engineering 22 & 23, September,
TN, IND. 2016 (ICMETE-2016) 22-23 Sept. 2016 - SRM
[26] Swati Jadon et. Al.,Face Recognition Using SOM University, Modinagar, UP, IND. (Accepted)
Neural Network with DDCT Facial Feature Extraction [30] Mahendra Kumar et. Al., Image Fusion Based On
Techniques, 4th IEEE International Conference on Hybrid SPIHT and SOMA, IEEE International
Communication and Signal Processing (ICCSP 2015) Conference on Micro-Electronics and
02-04 April 2015 - Melmaruvathur, TN, IND. Telecommunication Engineering 22 & 23, September,
[27] Reema Jain et. Al., Digital Image Watermarking 2016 (ICMETE-2016) 22-23 Sept. 2016 - SRM
using Hybrid DWT - FFT Technique with Different University, Modinagar, UP, IND.
Attacks, 4th IEEE International Conference on
Communication and Signal Processing (ICCSP 2015)
02-04 April 2015 - Melmaruvathur, TN, IND.

TABLE 2
COMPARISON AT EPOCH=1000

Technique Time Test Img 1 Test Img 2 Test Img 3 Test Img 4 Test Img 5 Test Img 6 Remarks
(Speed)

SED Training 444.9407 444.9407 444.9407 444.9407 444.9407 444.9407


Time

Execution 583.2211 618.3278 656.7674 690.6356 719.1419 742.9051 High speed


Time
DCT Training 584.3152 584.3152 584.3152 584.3152 584.3152 584.3152
Time

Execution 632.9374 696.4970 726.4075 751.5318 781.8307 804.9730 Low speed


Time
DWT Training 528.6507 528.6507 528.6507 528.6507 528.6507 528.6507
Time

Execution 582.0531 638.0103 671.4976 705.3821 731.0787 753.2664 Medi. Speed


Time
DDCT Training 534.2712 534.2712 534.2712 534.2712 534.2712 534.2712
Time
Execution 581.3175 610.4082 664.9835 702.5441 727.5838 791.8648 High Speed
Time
DCT-PT Training 174.4774 172.4774 172.4774 172.4774 172.4774 172.4774
Time
Execution 189.1488 203.6534 231.1831 286.4792 310.2474 331.3164 Very High Speed
Time

You might also like