You are on page 1of 6

International Conference on Communication and Signal Processing, April 3-5, 2018, India

Face Recognition using DRLBP and SIFT


Feature Extraction
M.Sushama and E.Rajinikanth


Abstract—The detection of human face from images plays a Finally, a local feature descriptor is computed at each key
vital role in Computer vision, cognitive science and Forensic point. This descriptor is based on the local image gradient,
Science. The various computational and mathematical models, for transformed according to the orientation of the key point
classifying face including Scale Invariant Feature Transform
to provide rotation and scale invariance. [d] Every feature is
(SIFT) and Dominant Rotated Local Binary Pattern (DRLBP)
have been proposed to yield better performance. This paper a vector of dimension 128 distinctively identifying the
proposes a novel method of classifying the human face using neighborhood around the key point.
Artificial Neural Network. This is done by pre-processing the face This paper proposes a user’s identity verification using
image at first and then extracting the face features using SIFT.
Dominant Rotate Local Binary Pattern and Scale Invariant
Then the detection of human faces is done using Back
Propagation Network (BPN). The process of combining SIFT and feature transformation. The software which is required for this
DRLBP perform better accuracy rather using separately. paper is MATLAB 7.0 FRAMEWORKS and PYTHON.
There are several ways which is mentioned by section wise.
Index Terms—Face Recognition, Dominant local binary Section I describes an Introduction of human face recognition.
pattern, Histogram and Feature Extraction, SIFT. Section II describes about many related works. Section III says
about the methodologies of preprocessing an image, DRLBP,
I. INTRODUCTION Histogram and GLCM Properties, Feature Extraction and
Back Propagation Network (BPN). Section IV describes the
A. HUMAN FACE RECOGNITION experimental Results and Output. Section V concludes the
paper with Conclusion and Future works.
Nowadays we’seen and come to know that aadhaar
II. RELATED WORK
identification problems for many people in government ,
public distribution systems and many other aspects. Got up The Conveyed Information is similar to some of the
with an inspiration to find out a solution with Face discussed methods are as follows.[1] Lei Xu, Lei Huang,
Recognition using DRLBP &SIFT Feature Extraction. Changping Lui, there is an another form of identifying the
effect of illumination problems that could change the basic
DRLBP &SIFT Feature Extraction procedure is an accurate appearance of an image in the human face recognition system.
and perfect solution for all identification problems. So, it many image processing method can toss out the low frequency
would be very useful in Banking Services, ATM Services, components of an image. Illumination can cause distortion of
Aadhaar linked programs, Police investigations and many processed images and perform poorly under normal lightening.
more service sectors. Firstly we choose a an appropriate images which can be
Scale Invariant Feature Transform (SIFT) are features named as raw vs processed images. Hence they lightened up
extracted from images to help in reliable matching with an idea that can dealed with illumination problems. So,
between different views of the same object. The extracted there is an adaptive method that can give robust performance
features are invariant to scale and orientation, and are after the combination of data base images.
highly distinctive of the image. [a] This computes the [2] Mehyar Emambakhsh, et all Currently the present
locations of potential interest points in the image by generation is facing the 3d technology issues. their
detecting the maxima and minima for a function of performance is increasing in day to day life. This imaging
Gaussian filters which is applied different scales all over systems has different noise components which can be deduced
the images and resampling those images. [b] These images by post processing methods in recognition systems. there is a
are discarded due to low contrast and an edge response best performance in denoising techniques of 3D image
points , then it is assigned to the dominant orientation.[c] processing methods. As per the denoising techniques an image
is re-alligned , resampled and later on the face is cropped
which results in the feature space as query and gallery. By this
M.Sushama and E.Rajinikanth are with the Sathyabama Institute of Science
and Technology, Chennai
way, the mask size of an image is larger compared to 3d facial
recognition system [3] Sanjay Thakre,In this there are two
cases where biometric features and finger print features are

978-1-5386-3521-6/18/$31.00 ©2018 IEEE

0994
compared to identify how many face images are attempted and that can be rectified in object tracking such as scale images ,
which has good performance. moving back ground and motion blur images. For improving
[4] an enhanced FR System using Fast Walsh Hadamard the detection rate in CCTV footages we can use the
Transform and Chiral Superimposition Technique as efficient heterogeneous training data and data augmentation. Moreover,
pre-processing methods and Dual sub-band based Discrete the growth of intrinsic camera parameters are increasing
Wavelet Transform(DWT) for feature extraction Niveditha G rapidly.
V, Sharmila B P , ] Present face recognition systems are facing
many problems which can be poses and illuminations. So they III. BLOCK DIAGRAM
tried to improve the performance of FR systems. three The block diagram of face detection is shown in Fig. 1.
techniques are used viz is Fast walsh hadamard transform,
chiral image superimposition and discrete wavelet transform. A. PREPROCESSING& NORMALISATION
All these techniques are used particular stages to improve the Normalisation of pre-processing takes place in the sample
overall performance of FR systems. values of the variable. A count value or data is set into an
[5] Method of pose neutralisation based on flipping an dimensions of certain image pixels. First one free space sample
image and then fusing with the original, and a novel method of is scaled and corrected up with original image. The sample
illumination equalisation(AIE) based on erosion and dilation
image has a pixel range of [ i, j ] where
S. Hitesh Babu, Shreyas H.R, FR systems has many challenges
in facing the effects of pose problems and illumination i contains ith no.of rows
problems under lightening conditions. This paper has a
combination of active illumination equalization,image j contains jth no.of columns
sharpening,standard deviation filtering,mirror image super
imposition and binary particle swarm optimization which is
used for the implementation of feature space in the optimal
feature subset[6] a Computer Vision application to enable
message composition by Head Tracking Laura Montanini,
Enea Cippitella, The FR system which is helpful for
identifying the robbers in an ATM securities and banking
purposes.The research work that is obtained from the previous
work is [6] the human face appearance which is utterly
change the pose, illumination and background of an image.
There variations in the Face recognition is difficult to
recognize when the more variations occurred . to solve these
type of difficulties we used two different techniques namely
EAFI and image folding. The selection of features in dwt and
dct transforms has high recognition rates. Fig. 1. Block diagram of face detection
[7] a real time transmission and display system of a spherical B. DOMINANT ROTATE LOCAL BINARY PATTERN
image stream from a robot Yoshito Okada, Takumi
It rotates the binary values of an image.Divide the window
Fujinami,the capable of identifying the head movements at
different angles and poses is obtained in this paper. Hence into cells (e.g. 16x16 pixels for each cell).For each pixel in a
there is a real time transmission occurs to display an image in cell, compare the pixel to each of its 8 neighbors (on its left-
the system. This system has an hard mount display with an imu top, left-middle, left-bottom, right-top, etc.). Suppose if the
so that it can perform the operations by varying fastly. all these pixel value is greater than the neighbour's value of 5 , then
operations can used for building an actual robot.[8] an the binary digit should betake as "1". Otherwise, it can be
effective robust method for detecting faces in video sequences taken as "0". This gives an 8-digit binary number which is
based on Haar features P.Ithava Rani, K. Ithaya
usually converted to decimal for convenience. Fig. 2
Muneeswaran,the face detection system plays a crucial role at
video based sequences in computer vision system. these represents the pixel range of the face image.
computer vision systems is used for video surveillance, face
recognition ,emotions recognition, movement recognition and
data base management. this paper has high detection rates in
video sequences which is capable of processing the images
and to detect the human faces.
[9] the use of heterogeneous training data and data
augmentation is explored to improve their detection rate in
challenging CCTV scenes A.Dimou, P.Medentziduo,
F.Alvarez Garcia,in this the CCTV footage is often used for Fig. 2. Represents the pixel range of the face image that includes the skin
length, size of the image.The strength of the skin varies from person to
object tracking due to zoom operations, low camera quality
person.
and diverse working environments. There are several problems

0995
C. SCALE INVARIANT FEATURE EXTRACTION G. CONTRAST
Scale Invariant Feature Transform (SIFT) are features This measures the local variations and texture of
extracted from images to help in reliable matching shadow depth in the gray level co-occurrence matrix.
between different views of the same object. The extracted
features are invariant to scale and orientation, and are
highly distinctive of the image. [a] This computes the H. CORRELATION
locations of potential interest points in the image by This measures the joint probability of the occurrence
detecting the maxima and minima for a function of of specified pixel pairs. Correlation may be a line of gray level
Gaussian filters which is applied at different scales all linear dependence between the pixels at the specified positions
over the images and re-sampling those images. [b] These relative to every alternative.
images are discarded due to low contrast and an edge
response points , then it is assigned to the dominant
orientation.[c] Finally , a local feature descriptor is I. HOMOGENITY
computed at each key point. This descriptor is based on Measures the closeness of the distribution of elements in the
the local image gradient, transformed according to the GLCM to the GLCM diagonal to the texture of an
orientation of the key point to provide rotation and scale
image.Angular Second Moment(ASM) could be a life of
invariance. [d] Every feature is a vector of dimension
128distinctively identifying the neighborhood around the homogeneity of a picture. A homogeneous scene can contain
key point. solely many gray levels giving with a GLCM co-occurance
matrix. However it comparatively has high values of p(i, j).
D. GLCM PROPERTIES AND CO-OCCURANCE MATRIX Thus, the addition of squares will be always high.
It is an image that ranges the corresponding pixel values into
the gray scale values or with other colours.It is defined for J. BACK PROPAGATION NETWORK
the specifying the spatial relationship with an pixel value i and
The performance of the Back Propagation network was
j .For an instance the position operator of the function (x,y)
that indicates (2,1) is applicable to an pixel value of an image evaluated in terms of training performance and classification
then it determines the "two down , one right". Thereby a accuracies. Back Propagation network gives fast and accurate
matrix is formed by an gray scale values which is known as classification and is a promising tool for classification of the
p*p occurrence matrix. In any image the spatial relationship tumors. The back-propagation learning rule can be used to
between two pixels i and j determines the sum of the both gray adjust the weights and biases of networks to minimize the sum
scale values of the image is equal to 1.
squared error of the network. Back propagation network
depends upon the feed forward neural network. Feed forward
E. ENERGY neural network is the simplest and easiest way in performing
the operations. It is based on the single layered neural network,
It is defined by the sum of squared elements in an rectangular hence it is known as artificial neural network. The following
image which had their own uniformity.
steps which is needed to perform in the artificial neural
network is mentioned below. They are as follows:
1. Activation of the input layer.
2
(1) 2. Transforming the information to the hidden
layer.
F. ENTROPY 3. weights of the output layer.
It gives a measure of complexity of the image. Complex The main goal of the back propagation network is supervised
textures tend to have higher entropy. this is often a live of learning algorithm. It used to map a set of input values to the
randomness, having its highest price once the weather of C output values. It is to train the multi layered neural network
area unit all equal. within the case of a game board, the which has an appropriate arbitrary values of input and output.
entropy would be low. This is function is to design values of one or more variable
into the real value or function. This can be also known an error
function or loss function. In the back propagation network,
error function finds the difference between the present input to
thee expected output.
(2) This network has certain modes of learning methods. They are
Where, stochastic and batch.

p(i , j) is the co occurrence matrix.

0996
K. Stochastic
In stochastic learning , each input has a weight
adjustment. So that every input should depend upon the
initialisation of weights. This stochastic produces more noise
using the local gradient process. Hence this calculates the
network points which can deduce the local minima.

L. Batch
In batch learning, weights are adjusted based on the
input values. In this the performance yields more faster and
more stable network. Thus it contains mini number of samples Fig. 5. shows that dominant rotate local binary pattern is applied for the input
and rest other samples are places in a batch from the data set. image

Limitations Here the input binary image is converted into gray scale
image. So, that the camera can identify the image very easily.
The limitations of back propagation network has no Fig. 6, 7, 8, 9 and 10 respresents the dominant rotate local
assurance to find the global minima of error function. Back binary pattern, gray scale image of left eye, right eye, nose and
propagation does not need the normalisation of input vectors mouth respectively. Fig. 11 detects the face image parts and
but normalisation has high performance. Fig. 12 identifies the particular location points. Fig. 13 and
Fig. 14 shows the aunthentication person and face recoginition
DRLBP ALGORITHM
performance analysis.
INPUT AND OUTPUT IMAGES
The input image, face detection and dominant rotate local
binary pattern are shown in Fig. 3, Fig. 4 and Fig. 5
respectively.

Fig. 3. InputImage Fig. 6. Represents the sinusoidal pulsating signal of dominant rotate local
binary pattern

Fig. 7. Represents the gray scale image of left eye

Fig. 4. Face detection

The above image shows that the image is scanned and the face
is cropped with an rectangular box. It descibes that the image
is in binary image.there by it can detect the face of the image. Fig. 8. Represents the gray scale image of right eye

0997
Fig. 9. Shows that the gray scale image of nose

Fig. 14. Shows the Performance Analysis of Face Recognition. Here the x-
axis is represented as number of samples are taken and y-axis represents the
face recognition accuracy
Fig. 10. Represents the gray scale image of mouth

IV. EXPERIMENTAL RESULTS


Hence, In this paper we came into an approach for Face
recognition using DRLBP&SIFT Feature Extraction.
Furthermore, the evaluation is done by using some of
parameters like accuracy, sensitivity and specificity which
could show best performance of the output shown in Fig. 15.

Fig. 11. Detecting the parts of the face image

Fig. 15. Parameters of the face recognition

A. COMPARISION

S.No Parameters Existing system Proposed


system

1. Accuracy 48% 75%

Fig. 12. Identifying the points of particular location B. PERFORMANCE

Serial Parameters Percentage(%)


no

1. Sensitivity 86%
Fig. 13. Shows the authentication person
2. Specificity 60%
For Example If any image is not identified from the date base
3. Accuracy 75%
then it will show that the person is an unauthenticated person.

0998
V. CONCLUSION AND FUTURE WORK [2] Mehyar Emambakhsh, Jiangning Gao, Adrian N Evans,” An
In this paper we came into an approach for Face Evaluation of Denoising Algorithms for 3D Face Recognition”
recognition using DRLBP &SIFT Feature Extraction. This can in International Conference on Imaging for Crime Detection and
be used in Banking Purposes, ATM Securities and many more Prevention, 2013.
Home applicable services. This idea has come through the [3] Sanjay Thakre,”Secure Reliable Multi-model Biometric
criminal detection that could help policemen to identify the Fingerprint and Face Recogition” in International Conference on
face of the criminal. Lamiaa A.Elrefaei et all has proposed a Computer Communication and Informatics,2017.
paper on real time face recognition on client-server based [4] Niveditha G V, Sharmila B P,” DWT-based Face Recognition
video sequences. this paper is used to find number of true using Fast Walsh Hadamard Transform and Chiral image
faces and number of false faces are identified. Hence, their Superimposition as preprocessing techniques”, in International
accuracy for finding the images is 48%. We proposed a Conference on Electronics and Communications, 2015.
framework on face recognition using dominant rotate local [5] S. Hitesh Babu, Shreyas H.R,”Face Recognition using Active
binary pattern and scale invariant feature extraction to detect Illumination Equalisation and Mirror Image Superimposition as
whether the person is an authenticate or an unauthenticated. Pre-processing techniques”, in International Conference on
Signal and Image Processing,2014.
Hence by using the parameters like sensitivity, specificity and
[6] LauraMontanini,Enea Cippitella, “Real Time Message
accuracy has high performance compared to the Viola-Jones
Composition Through Head Movements in Portable Android
algorithm. The Future scope of this paper can implement in
Devices”, IEEEInternational Conference on Consumer
video based features.
Electronics (ICCE) in 2014.
[7] Yoshito Okada, Takumi Fujinami,”Real-time Display of
ACKNOWLEDGMENT Onboard Spherical Images That Synchronizes Head Pose of
We are pleased to acknowledge our sincere gratitude to Operator for Robotic Search”, IEEE International Symposium
E.Rajinikanth, Asst.Professor, Department of ETCE, on Safety, Security, and Rescue Robotics (SSRR) in 2016.
Sathyabama Institute of Science and Technology, for guiding [8] P.Ithava Rani, K. Ithaya Muneeswaran ,“Robust Real Time Face
us in all the aspects we needed and his constant support, which detectionautomatically from video sequence based on Haar
lead us work towards the project effectively and progress features”, International Conference on Communication and
ahead through all the trouble shoots during the project. Network Technologies (ICCNT) in 2014.
[9] A.Dimou, P.Medentziduo, F.Alvarez Garcia,”Multi-target
REFERENCES detection in CCTV Footage for Tracking applications using
[1] Lei Xu, Lei Huang, Changping Lui, “Raw vs Processed: how to deep learning techniques”,IEEE in Spain Information
use the Raw and Processes images for Robust face Recogition Technology Institute in 2016.
under varying illumination”, in International conference on
Pattern Recognition, 2010.

0999

You might also like