Pose Invariant Face Recognition using Neuro-Fuzzy Approach

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661,p-ISSN: 2278-8727, Volume 17, Issue 3, Ver. II (May – Jun. 2015), PP 20-27
www.iosrjournals.org
DOI: 10.9790/0661-17322027 www.iosrjournals.org 20 | Page
Pose Invariant Face Recognition using Neuro-Fuzzy Approach
Reecha Sharma1
, M.S Patterh1
1
Department of ECE Punjabi University Patiala, Punjab, India
Abstract : In this paper a pose invariant face recognition using neuro-fuzzy approach is proposed. Here
adaptive neuro fuzzy interface system (ANFIS) classifier is used as neuro-fuzzy approach for pose invariant face
recognition. In the proposed approach the preprocessing of image is done by using adaptive median filter. It
removes the salt pepper noise from the original images. From these denoised images features are extracted.
Here Principal component analysis (PCA) is used for extracting the features of an image under test. Then
ANFIS classifier is used for face recognition. PCA calculate the principal components and are used by ANFIS
for further process. Here in this paper combination of PCA and ANFIS is represented as PCA+ANFIS. In the
paper standard ORL face database is used for experimental results. The performance PCA+ANFIS with
LDA+ANFIS and ICA+ANFIS is analyzed and compared. From experimental results it is shown that
PCA+ANFIS outperforms than other two approaches. PCA+ANFIS is also compared by existing feed forward
neural network (FFBNN) approach. The results show that proposed approach gives better outputs in terms of
accuracy, sensitivity and specificity.
Keywords - Feature extraction, image denoising, image recognition, neural networks and feedforward neural
network.
I. Introduction
These days security is one of the most common and important issue. Passwords are easily hacked by
hackers. So for more security unique passwords or something which only users have for authentication are
necessary. Human traits are unique and more secure for identification. These days human authentication is done
by biometrics. Many biometric traits are used for this purpose. One of the most common reasons of using
biometric traits is that they are unique and can’t be stolen, share or reproduced. Most commonly used traits are
palm, iris, finger prints, voice and face. All traits except face need presence of an individual for authentication.
But in case for face trait images of a person can be use for this purpose. Human can easily recognize faces but
by computer vision machines and pattern recognition it is challenging task. Face recognition can hurriedly and
correctly recognize target persons when the conditions are favorable. Researchers find many challenges in face
recognition. The most common challenges are pose, illumination, occulation, expression etc. In this proposed
research paper a pose invariant system for face recognition has been studied.
The rest of paper is organized as follows: Section II explains a brief knowledge about related work.
Section III explains schematic of proposed neuro-fuzzy approach for pose invariant face recognition. Section IV
gives experimental results and discussion and in section V conclusion and future scope of proposed work
explained.
II. Related Work
There are two main broad categories of face recognition. They are template based or feature based. In
first category i.e. template based full face is considered for recognition purpose. In second category i.e. feature
based only common features are used. These features may be eyes, nose, mouth and ear etc. combination of
feature is also used for improving the recognition rate [1]. The eigenfaces or Principle Components are the set of
characteristic features of a face obtained by decomposing a face. Eigenface gives training phase. For recognition
test image is placed into subspace called face space and recognition is done by comparing the test image with
trained database in face space [2]. Neural network architecture is used for pose invariant face recognition. PCA
is used for feature extraction and neural network is used for recognition [3]. It seems to be very difficult for
human to recognize faces correctly when the illumination varies severely, since the same person appears to be
very much different. Back propagation neural network is used for illumination invariant face recognition. The
features of an image are extracted in eigenspace. Then illumination direction specific back propagation neural
network trains the extracted feature and then testing is down [4]. Radial basis function (RBF) neural classifier
reduces the problem of small training set of high dimensional [5]. Non linear face images give 90% of
acceptance ratio when back propagation neural network is combined with PCA [6]. Combination of SOM and
convolution network gives fast and automatic system for face recognition as compare to Eigen faces [7]. When
PCA combined with feed forward neural network PCA-NN gives improved recognition as compare to
traditional PCA. The same improvement is shown in LDA-NN [8]. FFNN improves the performance of face

recognition as compare to Euclidean distance [9]. The novel architecture of the two-layer neural can recognize
human face with different views. It recognizes the identity of the person and the pose variation of a face at the
same time [10]. A combination of RBFNN and FHLA is used for face recognition. It provides faster training
with less number of neurons in hidden layer [11]. ANFIS and SVM fuse local and global features for face
verification. They fused together for better face recognition as compare to non adaptive and non fusion schemes
[12]. Linear Discriminant Analysis (LDA) is just like as eigenfaces of PCA. The only difference of LDA and
PCA is in the method of calculating the subspace. LDA maximize the ratio of the between class scatter matrix
and within class scatter matrix [13]. One hidden layer FFNN not only reduces the hidden units but also weights.
It gives better results as compare to BPNN [14]. Fusion of LDA and PCA give better results in face recognition
because they are intercorrelated to each other [15]. Independent component analysis (ICA) is a generality of
PCA. It separates the high order and second order moments. ICA is performed on images of face database by an
unsupervised learning algorithm derived from the optimal information transfer through sigmoidal neurons [16].
ICA, PCA and rough set theory combined to make a hybrid method for face recognition. ICA and PCA are used
for feature extraction and for recognition rough set rule based classifiers are used [17]. In this proposed paper
we are comparing PCA+ANFIS with LDA+ANFIS, ICA+ANFIS and also with FFBNN.
III. Schematic For Neuro-Fuzzy Approach
The proposed schematic for pose invariant face recognition using neuro-fuzzy approach is discussed in
this section. Here in the proposed approach PCA+ANFIS based pose invariant face recognition is done.
Proposed schematic for pose invariant face recognition using neuro-fuzzy approach have following steps.
1. The original images are taken from some standard face database. In our proposed schematic ORL face
database is used. The ORL face database images are grouped as test images and training images.
2. These images are preprocessed. Here hybrid filter called adaptive median filter is used for removing salt
and pepper noise from the original images.
3. These denoised images are given to PCA for feature extraction. Here principal components are calculated
and saved for further use.
4. These principal components are used by ANFIS a neuro-fuzzy classifier. The proposed ANFIS have five
layers. This is used for classification of images.
5. Here classified images are compared with predefined threshold value. If it greater than threshold value then
it is a recognized face otherwise not.
These steps are explained in detail in further sub sections.
3.1 Adaptive median filter
The adaptive median filter is used for preprocessor for noise removal. The noised images ),( srfd from ORL
face database are given as input to the adaptive median filter. These images are affected by the impulse and salt-
pepper noise. The output of adaptive median filter is noise free image. Fig.1 shows some samples of the noised
images from ORL face database. Fig.2 shows noise free output of adaptive median filter.
Figure 1: Sample of noised images from the ORL face database.
Figure 2: Denoised images after adaptive median filtering.

The working of adaptive median filter is explained in following steps:
Step 1: Initialize the window w size zw .
Step 2: Check if the center pixel ),( srpcen within w is noisy. If the pixel ),( srpcen is noisy go to step 3.
Otherwise slide the window to the next pixel and repeat step 1.
Step 3: Arrange all pixels within the window w in an ascending order. Then calculate minimum ( ),(min srp ),
median ( ),( srpmed ), and maximum ( ),(max srp ) values.
Step 4: Calculate if ),( srpmed is noisy,
(i.e.) ),(),(),( maxmin srpsrpsrp med  (1)
If the median value falls in the range of the minimum and maximum, it means that pixel is not a noisy pixel.
Then go to step 5. Otherwise ),( srpmed is a noisy pixel and then go to step 6.
Step 5: Here centre pixel of output image is replaced with ),( srpmed and go to step 8.
Step 6: In the step check whether all other pixels are noisy. If pixels are noisy then expend the window size by 2
and go to step 3. If not, go to step 7.
Step 7: Center pixel of noisy image is replaced with the noise free pixel which is the closest one of the median
pixel ),( srpmed .
Step 8: Reset window zw size. Center of window is changed to next pixel.
Step 9: Repeat all the steps until all pixels of an image are processed.
The table 1 shows the experimental results of noise removal filters. Average filter and Gaussian filter
are traditional filters used for noise removal but here in this work we are using adaptive median filter for noise
removal. The adaptive median filter is performed in spatial domain to find which pixels of an image are affected
by salt-pepper noise or impulse noise. In this each pixel is compared by neighbor pixels. The size of
neighborhood and threshold are adjustable. A pixel that is different from other neighborhood pixels is
considered as noise. These detected noise are replaced by median filter.
The main advantages of using adaptive median filter are as follow:
1. Remove impulse noise.
2. Smoothing of other noises
3. Reduces distortion of excessive thinning or thickening of boundaries.
In table 1, peak signal to noise ratio (PSNR) are calculated of some sample images. It is seen that
PSNR of adaptive median filter is better as compare to average filter and Gaussian filter. Here only few sample
images are taken for comparison.
Table 1 PSNR performance of Adaptive median filter, Average and Gaussian filter.
Images
PSNR
Adaptive
Median Filter
(in dB)
Average
Filter(in dB)
Gaussian
Filter (in
dB)
1 38.64 28.37 26.53
2 33.977 26.28 24.41
3 35.1861 26.81 26.02
4 34.54 26.19 25.52
5 33.96 26.68 25.08

Figure 3: Illustration of PSNR of sample images for adaptive median filter, average filter and Gaussian filter.
Fig.3 shows the illustration of PSNR of sample images for adaptive median filter, average filter and
Gaussian filter. The five sample images are taken to represent PSNR graphically. Adaptive median filter shows
highs peak.
3.2 Principal components calculation using PCA
The noise free images which are obtained from adaptive median filter as output are used for feature
extraction. Here feature extraction is done by PCA. The principal components are calculated here. The common
steps for PCA are given below:
1. Taking the whole dataset ignoring the class labels.
2. Computing the d-dimensional mean vector.
3. Compute the covariance matrix.
4. Computing eigenvectors and corresponding eigenvalues.
5. Sorting the eigenvectors by decreasing eigenvalues
6. Transforming the samples onto the new subspace
7. The Eigenvectors with highest eigenvalues are principal components of an image.
pca(x1), pca(x2), pca(x3)…………… pca(xn) are principal components obtained from the PCA process.
These principal components are then passed into neuro-fuzzy based ANFIS classifier for classification process.
3.3 Neuro-fuzzy ANFIS classifier
The principal components pca(x1), pca(x2), pca(x3)…………… pca(xn) calculated from the PCA are
classified using neuro-fuzzy ANFIS classifier. The architecture of proposed neuro-fuzzy ANFIS consists of five
layers of nodes. From these five layers, the first and the fourth layers have adaptive nodes whereas the second,
third and fifth layers possess fixed nodes. The architecture of the ANFIS for pose invariant face recognition is
given in fig.4
Figure 4: Architecture of ANFIS for pose invariant face recognition.

The learning process of ANFIS is carried out on the extracted PCA features such as Eigen vectors. The
Rule basis of the ANFIS is as follow:
If pca(x1) is Ai, pca(x1) is Bi and pca(xn) is Ci then
iniiii fxpcacxpcabxpcaaRules  )()()( 21 (2)
Where pca(x1), pca(x2), pca(x3)…………… pca(xn) are the inputs. iA iB & iC are the fuzzy sets,
iRules is the output within the fuzzy region specified by the fuzzy rule, ia , ib , ic and if are the design
parameters that are determined by the training process.
Layer 1: Every node i in this layer is a square node with a node function.
))(()),(()),(( ,12,11,1 nCiBiAi xpcaxpcaOxpcaO ii
  (3)
Usually
))(( 1xpcaiA
,
)(( 2xpcaiB
,
))(( nC xpcai

are chosen to be bell-shaped with maximum equal
to 1 and minimum equal to 0 and are defined as
i
ii
q
i
i
nC
BA
pca
ox
xpca
xpcaxpca













 




2
21
1
1
))((
))(())((


(4)
Where iii qpcao ,, are the parameter set. These are main parameters in this layer.
Layer-2: Layer 2 has circle nodes labeled by П. It multiplies the incoming signals and sends the product
output. For instance,
2,1
)),(())(())(( 21
,2



i
xpcaxpcaxpca
wtO
nCBA
ii
iii
 (5)
Each node output represents the firing strength of a rule.
Layer-3: Every node in this layer is a circle node labeled N . The
th
i node calculates the ratio of the
th
i rules
firing strength to the sum of all rule’s firing strengths:
2,1),/( 21,3  iwtwtwtwtO iii
(6)
Layer-4: Every node i in this layer is a square node with a node function
2,1.,4  iRuleswtO iii
(7)
Where iwt is the output of layer 3 and ia , ib , ic , if are the parameter set. Parameters in this layer will be
referred to as consequent parameters.
Layer-5: The single node in this layer is a circle node labeled  that computes the overall output as the
summation of all incoming signals:


 
i i
i ii
i
i
ii
wt
Ruleswt
RuleswtO ,5
(8)
21
2211
wtwt
RuleswtRuleswt
z


 (9)
21 RuleswtRuleswtZ  (10)

Then the predefined threshold value and the result of the neural network  Z is compared which is given in
Eq. (11).








Zrecognizednot
Zrecognized
result
,
,,
(11)
The neural network output Z greater than the threshold value ω means, the given input image is
recognized and Z less than the threshold value ω mean image is not recognized. Thus the ANFIS is well trained
using the score value obtained from PCA. The performance of the well trained ANFIS is tested by giving more
number of different pose images.
IV. Experimental Results And Discussions
The proposed schematic for pose invariant face recognition using neuro-fuzzy is implemented in
MATLAB. Here we compare the performance of filters for better noise removal. A sample image is denoised by
three filters. These are adaptive median filter, average filter and Gaussian filter. The results on ORL face
database shows that adaptive median filter gives high PSNR as compare to other two. Adaptive median filter
gives 38.64005 dB PSNR and on the other hand average filter gives 28.37dB and Gaussian filter gives 26.35dB.
Here results of PSNR shows that choice of adaptive median filter for noise removal is better as compare to other
filters.
Accordingly the denoised images acquired from the adaptive median filter are used to calculate the
principal components utilizing the PCA based calculation. The principal components in this way acquired from
the PCA are given as the input to the ANFIS classifier. More number of face images is used to analyze the
performance of the proposed face recognition system using different statistical performance measures.
The face images from ORL database are utilized to analyze the performance of neuro-fuzzy based
PCA+ANFIS approach with the ICA+ANFIS and LDA+ANFIS approach. The comparison results of the
proposed PCA+ANFIS, ICA+AFIS and LDA+ANFIS approach are shown in the table 1.
Table 2: Demonstrate the Performance comparison of the proposed neuro-fuzzy based PCA+ANFIS,
ICA+ANFIS and LDA+ANFIS.
Measures Proposed
PCA+ANFIS
(%)
ICA+ANFIS
(%)
LDA+ANFIS
(%)
Accuracy 96.66 71.3 68
Sensitivity 97.29 72.8 64.83
Specificity 96.05 71.2 72.88
In table 2 the accuracy of the proposed PCA+ANFIS approach is 96.66% but the ICA+ANFIS and
LDA+ANFIS approaches have offer only 71.3%, 68% of accuracy respectively. Similarly the sensitivity and
specificity of the proposed PCA+ANFIS approach is 97.29% and 96.05% but the ICA+ANFIS and
LDA+ANFIS approach give 72.8%, 64.83% of sensitivity and 71.2%, 72.88% of specificity respectively. Hence
from the table it can be seen that proposed approach recognizes the image more accurately.
Moreover proposed PCA+ANFIS is also compared with the existing FFBNN technique in terms of
sensitivity, specificity and accuracy and many more measures. In this approach the measurements are taken in
terms of true positive (TP) is correctly identified images by the approaches used for comparison, true negative
(TN) is correctly rejected images, false positive (FP) is incorrectly identify images and false negative (FN) is
incorrectly rejected images.
Here accuracy as also called true positive rate (TPR) and is given by
(12)
Specificity or true negative rate (TNR) is given by
(13)
False positive rate (FPR) is given by
(14)
Positive predictive value (PPV) is given by
(15)
Negative predictive value (NPV) is given by
(16)

The results are shown below in table 3.
Table 3: Illustrates the performance measures of the proposed PCA+ANFIS approach and the existing FFBNN
approach in terms of accuracy, sensitivity and specificity
Measures Proposed
PCA+ANFIS
Existing
FFBNN
Accuracy 0.9666 0.8666
Sensitivity 0.9729 0.8481
Specificity 0.9605 0.8873
FPR 0.0394 0.1126
PPV 0.96 0.8933
NPV 0.9733 0.84
FDR 0.04 0.106
MCC 0.9334 0.7343
From the table 3 it can be seen that the proposed PCA+ANFIS has given accuracy of 0.9666 but the
existing FFBNN has given accuracy of only 0.8666. Similarly the sensitivity and the specificity of our proposed
method are higher than the existing FFBNN. Fig.8 shows the illustrations of all measurement of PCA+ANFIS
and existing FFBNN.
Figure 8: Illustration of comparison of accuracies of proposed neuro-fuzzy based PCA+ANFIS approach with
ICA+ANFIS and LDA+ANFIS.
V. Conclusions
In this paper pose invariant face recognition using neuro-fuzzy approach is proposed. First the images
under test are denoised by using adaptive median filter and its performance is compared with average filter and
Gaussian filter. From the comparative result it has been found that adaptive median filter performs better as
compared to Average and Gaussian filter. PCA is used for feature extraction and ANFIS is used for face
recognition. The performance of the proposed PCA+ANFIS is compared with ICA+ANFIS and LDA+ANFIS.
From the comparative results it has been found that PCA+ANFIS perform better than ICA+ANFIS and
LDA+ANFIS. For example the proposed PCA+ANFIS give accuracy of 96.66% as compared to ICA+ANFIS
which gives 71.3% and LDA+ANFIS which gives 68%. Proposed PCA+ANFIS technique also performs better
than FFBNN. It has been concluded that PCA+ANFIS set up can be used for face recognition with better
accuracy. The proposed technique can further improve in future for other parameters like illumination,
occulation or age.
Acknowledgements
Author would like to express greatest gratitude to Dr. M.S Patterh Department of ECE, Punjabi
University for his continuous support for the paper from initial advice & contacts in the early stages of
conceptual inception & through ongoing advice & encouragement to this day. Author would also like to thank
the anonymous reviewers for their constructive comments. Author would like to thank AT&T Laboratories
Cambridge for providing the ORL face database.

References
[1]. Brunelli Roberto And Tomaso Poggio, "Face Recognition: Features Versus Templates," IEEE Transactions On Pattern Analysis
And Machine Intelligence, Vol. 15, No. 10, Pp. 1042-1052, 1993.
[2]. Turk Matthew And Alex Pentland, "Eigenface For Recognition," Journal Of Cognitive Neuroscience, Vol. 3, No. 1, Pp. 71-86,
1991.
[3]. Huang, Fu Jie, Zhihua Zhou, Hong-Jiang Zhang, And Tsuhan Chen, "Pose Invariant Face Recognition," In Fourth IEEE
International Conference On Automatic Face And Gesture Recognition, 2000, Pp. 245-250.
[4]. Wu-Jun Li, Chong-Jun Wan, Dian-Xiang Xu, And Shi-Fu Chen, "Illumination Invariant Face Recognition Based On Neural
Network Ensemble," In Proceedings Of The 16th IEEE International Conference On Tools With Artificial Intelligence (ICTAI
2004), 2004.
[5]. Ermeng Joo, Wu Shiqian, Lu Juwei, And Toh Hock Lye, "Face Recognition With Radial Basis Function (RBF) Neural Networks,"
Neural Networks, IEEE Transactions On, Vol. 13, No. 3, Pp. 697-710, 2002.
[6]. Mohammod Abul Kashem, Md. Nasim Akhter, Shamim Ahmed, And Md. Mahbub Alam, "Face Recognition System Based On
Principal Component Analysis (PCA) And Back Propagation Neural Network (BPNN) ," International Journal Of Scientific &
Engineering Research Volume 2, Issue 6, June-2011, Vol. 2, No. 6, Pp. 1-10, June 2011.
[7]. Lawrence Steve, Giles C Lee, Tsoi Ah Chung, And Back Andrew D, "Face Recognition: A Convolutional Neural-Network
Approach," Neural Networks, IEEE Transactions On, Vol. 8, No. 1, Pp. 98-113, 1997.
[8]. Alaa Eleyan And Hasan Demirel, "PCA And LDA Based Face Recognition Using Feedforward Neural Network Classifier,"
Multimedia Content Representation, Classification And Security, Pp. 199-206, 2006.
[9]. Alaa And Demirel, Hasan Eleyan, "Face Recognition System Based On PCA And Feedforward Neural Networks," In
Computational Intelligence And Bioinspired Systems, Pp. 935--942, 2005.
[10]. Huang Fu Jie, Zhihua Zhou, Hong-Jiang Zhang, And Tsuhan Chen, "Pose Invariant Face Recognition," In In Automatic Face And
Gesture Recognition, 2000. Proceedings. Fourth IEEE International Conference On, 2000, Pp. 245-250.
[11]. Haddadnia Javad, Faez Karim, And Ahmadi Majid, "A Fuzzy Hybrid Learning Algorithm For Radial Basis Function Neural
Network With Application In Human Face Recognition," Pattern Recognition, Vol. 36, No. 5, Pp. 1187-1202, 2003.
[12]. Fang, Yachun, Tieniu Tan, And Yunhong Wang, "Fusion Of Global And Local Features For Face Verification," In Pattern
Recognition, 2002. Proceedings. 16th International Conference On, Vol. 2, 2002, Pp. 382--385.
[13]. Eleyan Alaa And Demirel Hasan, Pca And LDA Based Neural Networks For Human Face Recognition.: INTECH Open Access
Publisher, 2007.
[14]. Ma Liying And Khorasani Khashayar, "Facial Expression Recognition Using Constructive Feedforward Neural Networks,"
Systems, Man, And Cybernetics, Part B: Cybernetics, IEEE Transactions On, Vol. 34, No. 3, Pp. 1588--1595, 2004.
[15]. Marcialis Gian Luca And Fabio Roli, "Fusion Of LDA And PCA For Face Verification," In Biometric Authentication, Pp. 30-37,
2002.
[16]. Bartlett Marian Stewart, "Independent Component Representations For Face Recognition," In Face Image Analysis By
Unsupervised Learning Springer US, Pp. 39-67, 2001.
[17]. Swiniarski, Roman W, And Andrzej Skowron, "Independent Component Analysis, Principal Component Analysis And Rough Sets
In Face Recognition," Transactions On Rough Sets I. Springer Berlin Heidelberg, Pp. 392-404, 2004.

Pose Invariant Face Recognition using Neuro-Fuzzy Approach

More Related Content

What's hot (18)

Similar to Pose Invariant Face Recognition using Neuro-Fuzzy Approach (20)

More from iosrjce (20)

Recently uploaded (20)

Pose Invariant Face Recognition using Neuro-Fuzzy Approach