International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 10

ISSN: 2321-8169
5884 - 5889

______________________________________________________________________________________

Face Recognition using Local Patterns

Mukesh D. Rinwa
Department of Information Technology
Amrutvahini College of Engineering (AVCOE)
Sangamner, Ahmednagar Dist.,India
e-mail: mukesh_rinwa@yahoo.com

Prof. Bharat S. Borkar

Department of Information Technology
Amrutvahini College of Engineering (AVCOE)
Sangamner, Ahmednagar Dist.,India
e-mail: borkar.bharat@gmail.com

Abstract Deriving an effective face representation is very essential task for automatic face recognition application. In this paper we used a
feature descriptor called the Local Directional Number Pattern (LDN), which allows individuals face recognition under different lightnings,
pose and expressions. Face recognition deals with different challenging problems in the field of image analysis and human computer interface.
To deal with attention in our proposed work we use local patterns, a local directional number pattern (LDN) method, a six bit compact code for
face recognition and understanding. By using LDN method we encode the directional information of the face images by convolving the face
image with the compass mask. This compass mask extracts the edge response values in eight directions in the neighborhood. For each pixel we
get the maximum and the minimum directional values which generate a LDN code i.e. generating an LDN image. Later LDN image is divided
into number of blocks for each block histogram is computed and finally adds these histogram from each block to form the feature vector which
acts as face descriptor to represent the face images. We perform different experiments under various illumination, pose and expression
conditions.
Keywords- Biometrics, face recognition, feature extraction, feature vector, Local patterns, Local Binary Pattern , Local Directional Number
Pattern .

__________________________________________________*****_________________________________________________
I.

INTRODUCTION

In the Recent years, automated method of recognizing a

person is going to be more popular like by using biometric
system. A biometric has shown an increasing interest in
many fields such as surveillances, human computer
interface, and security identification. Biometric recognition
is an automated method of recognizing an individual by
means of comparing the feature vector derived from the
behavioral and the physiological distinctiveness. Previously
in traditional based recognition system used Password,
PINs, Identity cards, Badges etc. There are several
limitations of using traditional based system like passwords
are hardly to remembered or sometimes it may be forgotten.
Tokens may be forgotten, mis-placed, stolen or lost.
Whereas, in biometric systems it cannot be stolen or forged.
It does not require any remember of password or PIN.
Various types of biometric techniques are Face recognition,
Iris recognition, Palm recognition, Retina recognition, Voice
recognition, Gait recognition etc. Among them face
recognition system is most easily accepted by user due to
the use of very few inexpensive cameras and also doesnt
require active user participation [1], [2].
There are two types of approaches to recognize an
individual: geometric-feature-based and appearance-based
methods. In Geometric feature based method which encodes
the shape and the locations information of different types of
the facial components which are added to form the feature
vector that represents the face images. But this method is
graph based method usually requires the accurate and
reliable facial feature detection and tracking system which is
very difficult in different environmental conditions. And the
appearance based methods used either holistic features or
local features to extract the different features from the face
images. To create the holistic features use image filters on
the whole face and to create local features use image filters
on some specific face region to extract the appearance
changes in the facial images. But the performance of the

holistic approaches degrades in environmental variations

[3]. There are various techniques used for the holistic
approach like eigenfaces, fisherfaces etc. and the techniques
used for the local approaches are Gabor features, Elastic
Bunch Graph Matching (EBGM), Local Binary Pattern
(LBP) etc among them LBP achieved better performance
its a 3*3 matrix which generates the LBP code by
computing the referenced or centre pixel with the
neighboring pixels. But this method works better in the
monotonic illumination changes but however the
performance degrades in the highly environmental
variations like illumination, random noise etc to overcome
these problems Local Ternary Patten (LTP), Local
Derivative Pattern(LDP) and Local Directional Pattern
(LDiP) was developed. The last method used directional
information instead of intensity to overcome the noise and
illumination variation problems. But these methods still lag
to extract the discriminated information to overcome this
Local Tetra pattern(LTrP) was developed which extracts the
information in the (n-1)th horizontal and the vertical
derivatives for efficient face recognition but this method
suffers from the high redundancy problems and feature
length increasing. To overcome the feature length problems
in our proposed work we use Local Directional Number
Pattern for efficient face recognition techniques, a six bit
code which extracts the directional information by
convolving compass mask with the input image and
obtaining the maximum and the minimum direction values
to generate the six bit micropatterns and corresponding an
LDN image is generated. Generated LDN image is later
divided into number of blocks to extract the histogram of the
each blocks and concatenating these histogram from each
block to form the feature vector which acts as the face
descriptor which later used to describe the face images.
The rest of this paper is discusses as follows. Section II
briefly discuss about what are the various local patterns used
for face recognition like LBP, LDP, LTP, LDiP, LTrP etc
which describe the properties of these methods for encoding
5884

IJRITCC | October 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 10

ISSN: 2321-8169
5884 - 5889

______________________________________________________________________________________
the micropatterns. In Section III discusses about the
proposed Local Directional Number Patterns (LDN) in
detail. Section IV discusses about the experimental results
and discussions. Finally conclusions are given in section V.
II.

LITERATURE SURVEY

discriminative to noise in uniform regions.LTP is a 3 coded

value 1,0, & -1. In LTP gray levels in a zone of width t
around ic are quantized to zero, one above this are quantized
to +1, and one below to it quantized to 1. One problem of
using this method is to set threshold t, which is not a simple
[8] [9].

In the proposed work we are working on the local pattern

approach for face recognition and understandings. This
section gives brief idea about what are the different local
pattern approaches used for face recognition like LBP, LDP,
LTP, LDiP, LTrP etc.
A. Local Binary Pattern
T. Ahonen et al and Zhang et al in [4], [5] defines LBP is a
simple 3*3 matrix which having one centre pixel and eight
surrounding neighbor pixels. The binary result is obtained
by computation of the reference pixel with all the eight
neighboring pixels. LBP is mainly used for texture analysis
of the image it works better in monotonic illumination
variations but however, it performance degrades with nonmonotonic illuminations variations like random noise etc.
and also LBP dont capture the dominant information in
large scale structure due to very less number of the
surrounding pixels [7], [8].

Figure 1.

The basic LBP operator [6].

B. Local Derivative Pattern

LBP method is not captured the dominant information in
very large scale structures due to sparse points of pixels to
overcome this limitation Zhang et al. in [6] proposed a new
descriptor called the Local Derivative Patterns(LDP), in
which the (n-1)th order derivative direction variations based
on a binary coding function, this LDP operator is 32 bit code
which encodes the higher-order derivative information
which contains more detailed discriminative features that the
first order LBP cannot obtained from an image.
In a general formulation, the nth- order LDP is a binary
string describing gradient trend changes in a local region of
directional (n-1)th-order derivative images 1 (Z) as
(Z0)= (1 (Z0), 1 (Z1)), (1 (Z0), 1 (Z2)),
, (1 (Z0), 1 (Z8))}
(1)
Where 1 (Z0) is the (n-1)th- order derivative in direction
at Z=Z0. (1 (Z0), 1 (Zi))} is defined as,
0, 1 Zi . 1 (Z0 ) > 0
(1 (Z0), 1 (Zi))}=
,
1, 1 Zi . 1 (Z0 ) 0
i=1,2,,8
(2)
The nth-order LDP is a local pattern string defined as,
LDPn(Z)= { (Z)|=00,450,900,1350}.
(3)
C. Local Ternary Pattern
LBP is sensitive to noise to overcomes this problem, Tan
and Triggs introduced a generalization of LBP called Local
Ternary Patterns(LTP) which is less sensitive and more

Figure 2. Illustration of basic LTP Operator [9].

D. Local Directional Pattern
LDiP descriptor considers the edge response values derived
from the kirsch gradient operator in eight directions around
a pixel, see Kirsch mask in fig.3. LDiP codes are generated
by setting the k most prominent directional bits to 1 and
remaining bits are set to 0. This method ignores the
distinction between the most prominent edge response
direction and the second most prominent one [10].
E. Local Tetra Pattern
Murala et al. in [11] found that the single high order
derivative direction (1D) relationship of the LDP can be
extended to the two higher order derivative direction (2D)
relationships in terms of the LTrP which adopts both the
horizontal and vertical high order derivative directions.
LTrP encodes with four distinct values by using 0 0 and 900
directions which extracts more detailed discriminative
information than the LDP which only considers 1D direction
with two distinct values. In order to encode the nth-order
LTrP, the (n-1) th-order derivatives along 0o and 900
direction.
And the direction of the center pixel can be calculated as,
1
1, 010 gc 0 and 90
0 (gc) 0
1

. (gc)=

1
2 010 gc < 0 90
0 (gc) 0
1
3 010 gc < 0 90
0 gc < 0

(4)

1
4 010 gc 0 and 90
0 (gc) < 0

The second-order LTrP2(gc) is defined as,

1
1
1
1
LTrP2(gc)={f3(
. (gc), . (g1)), f3( . (gc), . (g2)),,
1
1
f3( . (gc), . (gP))}|P=8
(5)
1
1
0,

.
.
1
1
f3(
(6)
. (gc), . (gP))=
1

. ,
Therefore, the generalized formulation for the nth-order
LTrP can be defined by using (n-1)th-order derivatives in
horizontal and vertical directions 1 (gP)|=00,900 as,
1
1
1
1
LTrPn(gc)={f3(
. (gc), . (g1)), f3( . (gc), . (g2)),,
1
1
f3( . (gc), . (gP))}|P=8
(7)
from (5) and (6), we get 8-bit tetra pattern for each center
pixel. After that separate all the patterns into four parts
based on the direction of center pixel. Finally, the tetra
patterns for each directions are converted to three binary
patterns.
5885

IJRITCC | October 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 10

ISSN: 2321-8169
5884 - 5889

______________________________________________________________________________________
LTrP2|Direction=2,3,4= =1 2(1)
4 (LTrP2(gc))|Direction=2,3,4
2

4 (LTrP (gc))|Direction

(8)
1, 2 =
= =
0,

(9)
Where = 2,3,4
Similarly, the other three tetra patterns for remaining three
directions (parts) of the center pixels are converted to binary
patterns. Thus, we get 12(4 3) binary patterns. Therefore,
totally 4 3 8=96 bit LTrP pattern is generated which
increases the feature length and high redundancy problems.
III.

PROPOSED SYSTEM

In our proposed work we are going to use Local

Directional Number Pattern (LDN), a six bit binary code
which is assigned to each pixel of an input image which
represents the structure of the texture and its intensity
transitions. In the Previous work like in LBP which is very
sensitive to illuminations variations and noise to overcome
this problem, we create our pattern using compass mask for
computing the edge response of the neighborhood by taking
the top directional numbers, which are the most positive and
the negative directions of the edge responses. These positive
and the negative responses provide essential information of
the structure of the neighborhood as they give the direction
of the bright and the dark areas in the neighborhood. These
most positive and the negative direction informations
generates a six bit LDN code corresponding to it an LDN
image is generated. The LDN image is divided into number
of blocks, for each block an histogram is computed which
forms the feature vector after combining these histograms
for each blocks. This procedure is done both for the training
image and the test image, comparing the feature vector
using the chi-square distance formula the value which is
minimum will gives the recognized image. This method is
robust as this method gives different values for different
regions of the intensity variations for example like dark and
the bright regions [12].

Our Proposed System is divided into following modules:

Image Sensor: The Image sensor (i.e. Camera) is used to
capture the face image. As we are using the standard face
database as a gallery and probe images, the block Image
sensor is shown here only for the completion of the
system diagram.
Face Alignment: Pre-processing the data before extracting
the features from the data, its function is to eliminate the
noise from the data, and to resize the data so that all the data
to be compared should have same dimensions. The images
in the PIE database are cropped and resized using
MATLAB- Image Processing Toolbox.
Feature Extraction: Before extracting the features from the
face image. It is preprocessed using the Kirsch and then
Sobel compass masks separately. After convolving with
each mask a high pass filtered image is obtained. This image
contains most of the information related with edges. The
edge contains prominent information for the discrimination
between the objects. From this eight images, taking into the
directional information in the positive (max value) and
negative (min value) a LDN code for each pixel is
generated. Repeating this procedure for all pixels finally we
get an LDN image. This LDN image is then divided into
4x4 regions. The histogram of each region is calculated.
Hence we have 16 histograms for a LDN image. These
histograms represents feature vector for the given face
image. This face features are termed as face descriptor.
Each Object (Face) has its own face descriptor.
Classification: To recognize/classify the object, its face
descriptor is compared with the face descriptor of all objects
in the database. For this we are using Chi-Square
dissimilarity measure (distance formula) between the two
face descriptors. The object (face) which has smallest
(value) difference after comparison with the gallery images
(training database) is treated as recognized object. Such
algorithm is termed as Nearest Neighbor (NN) algorithm.

Figure 3 . Kirschs Template in eight directions [12].

In the our proposed work we use Kirsch compass mask and
the Sobel operator which extracts the edge response of the
face image as well as smoothing the face image from noise
and illumination variation to produce a robust code.

Figure 4. Sobel Template in eight directions.

Figure 5. Proposed Face Recognition Architecture Using

LDN.
Algorithm:
Input: Query image; Output: recognized image
1. Load the image, and convert it into grayscale
image.
2. Apply the compass mask with the original face
image to get the directional information.
3. Select the maximum and the minimum position
direction from the neighborhood of the face image.
5886

IJRITCC | October 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 10

ISSN: 2321-8169
5884 - 5889

______________________________________________________________________________________
4.

5.
6.
7.

8.

9.

Concatenate the maximum and the minimum

position direction to form the LDN code, and
corresponding LDN image is generated.
Divide the LDN image into number of blocks, for
each block calculates their histogram.
Concatenate the histograms from each block to
form the feature vector.
Similarly, calculate the histogram for training
image to form the feature matrix (this step requires
only once at the first time when the application is
initialized.)
Compare the feature vector of the query image with
the feature vector of training image using ChiSquare dissimilarity measure.
The value which is minimum indicates the
recognized image.

Mathematical Modeling
f(x,y) denotes the input image of an object, where x & y
represents the coordinates of an image and f(x,y) represents
intensity or gray level of an image. The input image may be
color or monochrome. If it is color image it is converted into
gray image first and then processed. Here in the proposed
work we use two different compass masks to generate the
LDN code which convolve with input image in 8 directions
to obtain the directional information.
The input image f(x,y) is processed with masks to get the
LDN code
LDN(x,y )= 8ix,y+jx,y
(10)
Where (x,y) is the central pixel of the neighborhood being
coded, ix,y is the directional number of the maximum
positive response, and jx,y is the directional number of the
minimum negative response.
The directional number for maximum & minimum response
is calculated by
i x,y = arg max{i(x,y) | 0 i 7}
(11)
i
j x,y = arg max{j(x,y) | 0 j 7}
(12)
j
Where, i is the convolution of the original image I.
The convolution is done by
i= I*Mi
(13)
Where I is the original image and Mi is ith mask.
The Kirsch compass mask used is as shown in figure 3. In
addition to it we are using Sobel mask which is calculated as
gradient space as per the equations given below.
The sobel mask is calculated by,
gx = df/dx = (z7+2z8+z9) - (z1+2z2+z3)
(14)
and,
gy = df/dy = (z3+2z6+z9) - (z1+2z4+z7)
(15)
The magnitude of the gradient is approximated by
M(x,y)= |gx| + |gy|
(16)

The histogram obtained such represent the feature vector

for that individual face. Hence for all faces in the database
we have this feature vector. These are then compared with
Chi-Square dissimilarity measure. The value which is
minimum indicates the recognized image.
2 (1 , 2 ) =

1 2

=1 +
1
2

(18)

Where F1 and F2 are two feature vectors.

Similar steps are taken with Sobel Operator.
IV.

RESULT ANALYSIS AND DISCUSSIONS

In the Proposed work we use LDN method for face

recognition by using Kirsch and the Sobel operator as the
compass mask. Here we perform different operations for
face recognition under different Illumination, Pose &
Expression conditions. In the previous system work is done
by using Matlab here we are doing by using Java by using
JAI(Java Advance Imaging) as API.
The various experiments are conducted on the different
databases as given in the tables I, II, & III.
As the number of objects taken are different from the
different database, the graph is shown separately for each
database.
Table No I. Performance Analysis of ORL Database
Experi
ment

Gallery
Images
(Training)

Probe
Images
(Testin
g)

Correctly
recognized
Images using

Recognition
Rate = No. of
images
recognized
correctly/Tot
al no. of
images in the
Probe set *
100 (%)

No.
of
Obj
ects

No. of
face
image/
object

2 face
images/
object

Kirsc
h
Com
pass
Mas
k

Sobel
Oper
ator

Kirsc
h
Com
pass
Mas
k

Sobel
Oper
ator

10

20

19

18

95

90

20

40

37

35

92.5

87.5

40

80

75

72

93.7

90

Finally, a histogram is calculated for each block of the LDN

image.

LH=
(17)
=1
Where, is the concatenation operation, and N is the
number of regions of the divided face.

Figure 6. Graph showing Recognition Rate in ORL

Database
5887

IJRITCC | October 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 10

ISSN: 2321-8169
5884 - 5889

______________________________________________________________________________________
Table No II. Performance Analysis of YALE Database
Experi
ment

Gallery
Images
(Training)

Probe
Images
(Testin
g)

Correctly
recognized
Images using

Recognition
Rate = No. of
images
recognized
correctly/Tot
al no. of
images in the
Probe set *
100 (%)

No.
of
Obj
ects

No. of
face
image/
object

2 face
images/
object

Kirsc
h
Com
pass
Mas
k

Sobel
Oper
ator

Kirsc
h
Com
pass
Mas
k

Sobel
Oper
ator

10

10

90

100

10

20

17

19

85

95

15

30

25

26

83.3

86.6

Figure 8. Graph showing Recognition Rate in PIE Database

Figure 9. GUI Shows Face is recognized under different

Pose and Expression Conditions.
V.
Figure 7. Graph showing Recognition Rate in YALE
Database
Table No III. Performance Analysis of PIE Database
Experi
ment

Gallery
Images
(Training)

Probe
Images
(Testin
g)

Correctly
recognized
Images using

Recognition
Rate = No. of
images
recognized
correctly/Tot
al no. of
images in the
Probe set *
100 (%)

No.
of
Obj
ects

3 face
images/
object

Kirsc
h
Com
pass
Mas
k

Sobel
Oper
ator

Kirsc
h
Com
pass
Mas
k

Sobel
Oper
ator

No. of
face
image/
object

15

12

12

80

80

24

21

20

87.5

83.3

10

30

25

22

83.3

73.3

CONCLUSION

Biometric technique is very useful compared to other

traditional techniques like remembering of Password and
PINs for individuals identity. In Biometric, Face
recognition system is more useful for individuals identity
due to the using of very few inexpensive cameras and
doesnt require user active participation or any health risks.
In the proposed system we use LDN method, which is six
bit binary code obtain by Convolving the kirschs as filter
along with Sobel operator with input image. This six bit
LDN code is compact as compared to other method like
LBP which is eight bit binary code. This coding scheme is
based on the directional information, instead of small piece
strings, which encodes the information of the neighborhood
in a more effective way.
ACKNOWLEDGMENT
This work is being implemented under the guidance of
Prof. Bharat S. Borkar. His active guidance and valuable
suggestion at every time encouraged me to carry out the
same.
REFERENCES
[1] S. Prabhakar, S. Pankanti, and A. K.
Jain,Biometric recognition: Security and privacy
concerns, IEEE Security Privacy, vol. 1, no. 2,
Mar. /Apr. 2003, pp. 33-42.
[2] Anil K. Jain , An Introduction to Biometric
Recognition, IEEE Transactions on Circuits And
5888

IJRITCC | October 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 10

ISSN: 2321-8169
5884 - 5889

______________________________________________________________________________________
Systems For Video Technology, Vol. 14, No. 1,
January 2004, pp. 4-20.
[3] S.Z., A.k.Jain, Y.L.,Tian et al., Facial Expression
analysis, in Handbook of face Recognition. New
York: Springer-Verlag,2005, Chap. 11.
[4] T.Ahonen, A.Hadid, and M.Pietikainen, Face
recognition with local binary patterns, in Proc.
Euro. Conf. Comput. Vis., 2004, pp. 469-481.
[5] T.Ahonen, A.Hadid, and M.Pietikainen, Face
Description with local binary patterns: Application
to Face Recognition, IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol 28,
No.12, Dec.2006, pp.2037-2041.
[6] B. Zhang, Y. Gao, S. Zhao, and J. Liu, Local
derivative pattern versus local binary pattern: Face
recognition with higher-order local pattern
descriptor, IEEE Trans. Image Process., vol. 19,
no. 2, Feb. 2010, pp. 533544.
[7] Bo Yang & Songcan Chen, A comparative study
on local binary pattern (LBP) based face
recognition: LBP histogram versus LBP image,
Elsevier, Neurocomputing, 3 April 2013, pp. 365379.
[8] Di Huang, Shan, Ardabilian, Wang, & Chen,
Local Binary Patterns and Its Application to
Facial Image Analysis: A Survey, IEEE
Transaction on Systems, Man, & Cybernetics, vol.
14, No. 6, Nov. 2011, pp. 765-781.
[9] Tan, and Bill Triggs, Enhanced Local Texture
Feature Sets for Face Recognition Under Difficult
Lighting Conditions, IEEE Transaction on Image
Processing, vol. 19, No. 6, June 2010, pp. 16351650.
[10] T.Jabid, M.H. Kabir, O.Chae, Robust Facial
Expression Recognition Based on Local
Directional Pattern, ETRI Journal, Volume 32,
Number 5, October 2010, pp. 784-794.
[11] S. Murala, R. P. Maheshwari, and R.
Balasubramanian, Local tetra patterns: A new
feature descriptor for content-based image
retrieval, IEEE Trans. Image Process., vol. 21, no.
5, May 2012, pp. 28742886.
[12] Adin Ramirez Rivera, Jorge Rojas Castillo, and
Oksam Chae, Local Directional Number Pattern
for Face Analysis: Face and Expression
Recognition, IEEE Transactions on image
processing, vol.22, No.5, May 2013, pp.1740-1752.

5889
IJRITCC | October 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________