Vol 05, Article 05363; May 2014

International Journal of Engineering Sciences Research-IJESR

http://ijesr.in/ ISSN: 2230-8504; e-ISSN-2230-8512


2010-2014 - IJESR
Indexing in Process - EMBASE, EmCARE, Electronics & Communication Abstracts, SCIRUS, SPARC, GOOGLE Database, EBSCO, NewJour, Worldcat,
DOAJ, and other major databases etc.,
1527
TRENDY REGULARIZED ROBUST CODING FOR FACE
RECOGNITION
V.EZHILYA
1

1
HOD, Department Of ECE, VSA GROUP OF INSTITUTIONS, Salem, TamilNadu, INDIA
1
ezhilyavenkat@gmail.com

ABSTRACT
In present, automatic face recognition of people is a challenging problem which has received much attention
during the recent years due to its many applications in different fields. Recently, sparse representation has been
attracting a lot of attention due to its great success in image processing. In SRC, all training samples without
disguise are used to compose an overcomplete dictionary, and the testing sample with disguise is represented
by the dictionary with a sparse coding coefficients plus an error. The coding residuals between the sample and
each class of training samples are measured and the minimum of them is the identified class to which the sample
belongs. It involve maximum likelihood estimation (MLE) process, which involves iteratively sparse
representation based coding (ISRC),but it does not provide robust solution for face recognition and also require
high computational cost. In this paper, we proposed an extensive experiments on representative face databases
demonstrate that a novel Regularized Robust Coding (RRC) is much more effective and efficient methods in
dealing with complicated variations such as face occlusion, corruption, lighting and expression changes, etc
with the help of maximum a posterior (MAP) principle which are using iteratively reweighted regularized robust
coding(IRC) algorithm.
Keywords: Face Recognition, ISRC, Maximum a posterior principle, Maximum likelihood estimation,
Occlusions, Regularized Robust Coding, Sparse representation.


1. INTRODUCTION
In the last few decades, face recognition has attracted more and more attention in the field of computer vision
and patter recognition. As one of most successful applications in biometrics, face recognition can be applied in
social robotics to fulfill the person identification task in a natural and noncontact way. A picture speaks more
than words. In todays world, images play an important role in real time operations. The field of image
processing has grown vigorously in recent times. Digital image processing deals with manipulation and analysis
of images by using computer algorithm, so as to improve pictorial information for better understanding and
clarity. This area is characterized by the need for extensive experimental work to establish the viability of
proposed solutions. Image processing involves the manipulation of images to extract information to emphasize
or de-emphasize certain aspects of the information, contained in the image or perform image analysis to extract
hidden information.

Recent advances in Precision Farming have resulted in significant improvements in agriculture by increasing
crop production, with good quality and low operating cost. In this context, image processing is used for a
number of application areas that include soil testing, good quality seed selection, identification of nutrient
deficiencies, monitoring and controlling of stress, weeds and diseases, assessment of crop status and yield.
In practice, face patterns are subject to changes in illumination, pose, facial expression, etc. Among them, face
recognition with the real disguise is a very important and hard problem to be solved. Therefore, robust
visionbased face recognition has been extensively studied by researchers from the area of computer vision,
robotics, artificial intelligence, etc. Generally, face image is stretched into a high dimensional face vector, then
feature extraction and dimensionality reduction algorithms can be applied in the face space, so that the
highdimensional face vector is transformed into a lowdimensional subspace. And in this face subspace
classification and identification task can be implemented.
Recently, sparse representation is introduced from compressive sensing theory into the field of pattern
recognition; the sparse representationbased classification (SRC) is a landmark algorithm for robust face
recognition, which can deal with face occlusion, corruption and real disguise. The basic idea of SRC is to
represent the query face image using a small number of atoms parsimoniously chosen out of an overcomplete
dictionary which consists of all training samples. The sparsity constraint of the coding coefficients is employed
to insure that only a few samples from the same class of the query face have distinct nonzero values, whereas
the coefficients of other samples are equal or close to zero.
The sparsity of the coding coefficient can be directly measured by l
0
norm, which counts the number of nonzero
entries in a vector. However, the l
0
norm minimization is an NPhard problem; therefore the l1-norm
minimization is widely employed instead of the above problem. It has been demonstrated that l
0
norm and
Vol 05, Article 05363; May 2014

International Journal of Engineering Sciences Research-IJESR
http://ijesr.in/ ISSN: 2230-8504; e-ISSN-2230-8512


2010-2014 - IJESR
Indexing in Process - EMBASE, EmCARE, Electronics & Communication Abstracts, SCIRUS, SPARC, GOOGLE Database, EBSCO, NewJour, Worldcat,
DOAJ, and other major databases etc.,
1528
l
1
norm minimizations are equivalent if the solution is sufficiently sparse. The representation fidelity of SRC is
measured by the l
2
norm of the coding residual, which actually assumes that the coding residual follows
Gaussian distribution. It may not be able to effectively describe the real model of the coding residual in practical
situation of face recognition, especially dealing with real face disguise. The robust sparse coding (RSC) seeks
for the MLE (maximum likelihood estimation) solution of the sparse coding problem, so that the distribution of
coding residual is more accurate than Gaussian or Laplacian and it is more robust to disguise than SRC.

2. RRC
In recent years, sparse representation has been attracting a lot of attention due to its great success in image
processing, and it has also been used for FR and texture classification.Based on the findings that natural images
can be generally coded by structural primitives that are qualitatively similar in form to simple cell receptive
fields, sparse coding represents a signal using a small number of atoms parsimoniously chosen out of an over
complete dictionary. The sparsity of the coding coefficient can be measured by l
0
-norm, which counts the
number of non-zero entries in a vector.Since the combinatorial l
0
-norm minimization is an NP-hard problem, l
1
-
norm minimization, as the closest convex function to l
0
-norm minimization, is widely employed in sparse
coding, and it has been shown that l
0
-norm and l
1
-norm minimizations are equivalent if the solution is
sufficiently sparse.In general, the sparse coding problem can be formulated as
Min

||||
1
s. t ||y-D||
2
(1)
where y is the given signal, D is the dictionary of coding atoms , is the coding vector of y over D, and > 0 is
a constant.
By coding a query image y as a sparse linear combination of all the training samples via Eq. (1), SRC classifies
y by evaluating which class could result in the minimal reconstruction error of it.However, it has been indicated
in that the success of SRC actually owes to its utilization of collaborative representation on the query image but
not the l
1
-norm sparsity constraint on coding coefficient.One interesting feature of SRC is its processing of face
occlusion and corruption. More specifically, it introduces an identity matrix I as a dictionary to code the outlier
pixels
min

|| [; ] ||
1
s. t. y = [D, I][ ; ] (2)

By solving Eq. (2), SRC shows good robustness to face occlusions such as block occlusion, pixel corruption and
disguise. It is not difficult to see that Eq. (2) is basically equivalent to min

||||
1
s .t. ||y-D ||
1
< . That is, it uses
l
1
-norm to model the coding residual y- D to gain certain robustness to outliers.The SRC has close relationship
to the nearest classifiers.SRC could be seen as a more general model than the previous nearest classifiers, and it
uses the samples from all classes to collaboratively represent the query sample to overcome the small-sample-
size problem in FR.In addition, different from the methods such as LBP, Probabilistic local approach which use
local region features, color features or gradient information to handle some special occlusion ,SRC shows
interesting results in dealing with occlusion by assuming a sparse coding residual, as in Eq. (2).There are many
following works to extend and improve SRC, such as feature-based SRC, SRC for face misalignment or pose
variation, regularized collaborative representation and SRC for continuous occlusion.

2.1 Modeling of RRC
The conventional sparse coding model in Eq.(1) is equivalent to the so-called LASSO problem:
min

|| y-D||
2
s. t. |||| (3)
Where >0 is a constant, y = [y
1
;y
2
;.;y
n
]
n
is the signal to be coded ,D = [d
1
,d
2
,,d
m
]
n m
is the
dictionary with column vector d
j
being its j
th
atom, and
m
is the vector of coding coefficients. In the
problem of FR, the atom d
j
can be simply set as the training face sample and hence the dictionary D can be the
whole training dataset.If we have the prior that the coding residual e = y - D follows Gaussian distribution, the
solution to Eq.(3) will be the MLE solution. If e follows Lapacian distribution, the l
1
-sparsity constrained MLE
solution will be,
min

||y D||
1
s. t. ||||
1
(4)
The above Eq.(4) is essentially another expression of Eq.(2) because they have the same Lagrangian
formulation: min

{||y D||
1
+||||
1
}.In practice, however, the Gaussian or Laplacian priors on e may be
invalid, especially when the face image y is occluded, corrupted, etc.Inspired by the robust regression theory, in
our previous work we proposed an MLE solution for robust face image representation. Rewrite D as D=[r
1
;r
2
;
,, ;r
n
],where r
i
is the i
th
row of D, and let e = y-D = [e
1
;e
2
; , , ; e
n
], where e
i
=y
i
r
i
, i= 1,2,..,n. Assume
that e
1
,e
2
,..,e
n
are independent and identically distributed and the PDF of
i
is f

(e
i
), where denotes the
unknown parameter set that characterizes the distribution, the so- called RSC was formulated as the following
l
1
-sparsity constrained MLE problem [ let (e)= -lnf(e)]
min

(y
i
-r
i
) s. t. ||||
1
(5)
Vol 05, Article 05363; May 2014

International Journal of Engineering Sciences Research-IJESR
http://ijesr.in/ ISSN: 2230-8504; e-ISSN-2230-8512


2010-2014 - IJESR
Indexing in Process - EMBASE, EmCARE, Electronics & Communication Abstracts, SCIRUS, SPARC, GOOGLE Database, EBSCO, NewJour, Worldcat,
DOAJ, and other major databases etc.,
1529
Like SRC, the above RSC model assumes that the coding coefficients are sparse and uses l
1
-norm to
characterize the sparsity. However, the l
1
-sparsity constraint makes the complexity of RSC high, and recently it
has been indicated that the l
1
-sparsity constraint on is not the key for the success of SRC. In this project, we
propose a more general model, namely RRC. The RRC can be much more efficient than RSC, while RSC is one
specific instantiation of the RRC model.Lets consider the face representation problem from a view-point of
Bayesian estimation, more specifically, the MAP estimation. By coding the query image y over a given s
dictionary D, the MAP estimation of the coding vector is = arg max

ln p (|y). Using the Bayesian formula,

we have
=argmax

{ln p(y|)+ln p()} (6)

Assuming that the elements e
i
of coding residual e= y D =[e
1
;e
2
; ;e
n
] are i.i.d. with PDF f

(e
i
),we have p
(y|) =

(y
i
r
i
). Meanwhile, assume that the elements
j
,j =1,2,,m, of the coding vector =
[
1
;
2
;,
m
] are i. i. d. with PDF f
o
(
j
), there is p() =
o
(
j
).The MAP estimation of in Eq.(6) is
= arg max

(y
i
r
i
) +
o
(
j
)} (7)
Letting

(e) = - ln f

(e) and
o
() = - ln f
o
(), Eq.(7) is converted into
=arg min

(y
i
-r
i
) +
0
(
j
) } (8)
We call the above model RRC because the fidelity term

(y
i
-r
i
) will be very robust to outliers,
while
0
(
j
) is the regularization term depending on the prior probability p().
It can be seen that
0
(
j
) becomes the l
1
- norm sparse constraint when
j
is Laplacian distributed, p()
= (-||
j
||
1
/

)/2

. For the problem of classification, it is desired that only the representation

coefficients associated with the dictionary atoms from the target class could have big absolute values. As we do
not know beforehand which class the query image belongs to, a reasonable prior can be that only a small percent
of representation coefficients have significant values. Therefore, we assume that the representation coefficient j
follows GGD. There is
f
o
(
j
) = exp {-(|
j
|/

}/(2

(1/)} (9)
Where denotes the gamma function.

For the representation residual, it is difficult to predefine the distribution due to the diversity of image
variations. In general, we assume that the unknown PDF f

(e) are symmetric, differentiable, and monotonic w.r.t

|e|, respectively. So

(e) has the following properties:(1)

(0) is the global minimal of

(x); (2) symmetry:

(x) =

(-x); and (3) monotonicity:

(x
1
)>

(x
2
) if |x
1
|>|x
2
|. Without loss of generality, we let

(0) =0.

The proposed RRC model in Eq. (8) has close relations to robust estimation, which also aims to eliminate the
effect of outliers. The robust estimation methods, e.g., Regression Diagnostics, M-estimator and Least Median
of squares are widely used in parameter estimation and has various applications in computer vision, such as
tracking , robust subspace learning and so on.
However, there are clear differences between the previous robust estimation methods and the proposed RRC.
Most of previous robust estimation methods regard the whole pieces of samples but not the elements of a sample
as inliers or outliers. Although the robust subspace learning method weights each pixel by the judgment of inlier
or outlier, it aims to learn robust principle components but not to solve the regularized coding coefficients of a
testing sample with outliers. Besides, the proposed RRC model is developed in order for classification tasks but
not regression.
Two key issues in solving the RRC model are how to determine the distribution

(or f

), and how to minimize

the energy functional. Simply tasking f

as Gaussian or Laplacian and taking f

as Laplacian, the RRC model

will degenerate to the conventional sparse coding problem in Eq.(3) or Eq.(4). However, such preset
distributions for f

have much bias and are not robust enough to outliers, and the Laplacian setting of f
o
makes
the minimization inefficient. In this paper, we allow f
o
to have a more flexible shape, which is adaptive to the
input query image y so that the system is more robust to outliers. To this end, we transform the minimization of
Eq.(8) into an iteratively reweighted regularized coding problem in order to obtain the approximated MAP
solution of RRC effectively and efficiently.

2.2 RRC via Iteratively Reweighting
Let F

(e)=

(e
i
).The Taylor expansion of F

(e) in the neighborhood of e

0
is

(e)=F

(e
0
)+ (e-e
0
)
T
F
'

(e
0
)+R
1
(e) (10)
Vol 05, Article 05363; May 2014

International Journal of Engineering Sciences Research-IJESR
http://ijesr.in/ ISSN: 2230-8504; e-ISSN-2230-8512


2010-2014 - IJESR
Indexing in Process - EMBASE, EmCARE, Electronics & Communication Abstracts, SCIRUS, SPARC, GOOGLE Database, EBSCO, NewJour, Worldcat,
DOAJ, and other major databases etc.,
1530
Where R
1
(e) is the high order residual, and F

(e) is the derivative of F

(e). Donote by

the derivative of

,
and there is F

(e
0
) = [

(e
0,1
);

(e
0,2
);;

(e
0,n
)], where e
0
,i is the i
th
element of e
0
. To make F

(e) strictly
convex for easier minimization, we approximate the residual term as R
1
(e)0.5(e-e
0
)
T
W(e-e
0
), where W is a
diagonal matrix for that the elements in e are independent and there is no cross term of e
i
and e
j
,ij, in F

(e).
Since F

(e) reaches its minimal value (i.e.,0) at e=0, we also require that its approximation

(e) reaches the

minimum at e=0. Letting

'(0)=0, we have the diagonal elements of W as:

W
i,i
=

(e
0,i
) =

'
(e
0,i
)/e
0,i
. (11)
According to the properties of

, we know that

(e
i
)will have the same sign as e
i
. So W
i,i
is a non-negative
scalar. Then F

(e) can be written as

'
(e)= ||W
1/2
e||
2
2
+b
e0
(12)
Where b
e0
=

(e
0,i
)-

(e
0,i
)e
0,i
/2] is a scalar constant determined by e
o
.Without considering the constant
be0, the RRC model in Eq.(8) could be approximate as,
=arg min

{ ||W
1/2
(y D)||
2
+
0
(
j
)} (13)
Certainly, Eq.(13) is a local approximation of Eq.(8) but it makes the minimization of RRC feasible via
iteratively reweighted l
2
-regularized coding, in which W is updated via Eq.(11). Now, the minimization of RRC
is turned to how to calculate the diagonal weight matrix W.

2.3 Weights W
The element W
i,i
, i.e.,

(e
i
), is the weight assigned to pixel i of query image y. Intuitively, in FR the outlier
pixels should have small weights to reduce their effect on coding y over D. Since the dictionary D, composed of
non-occluded/non-corrupted training face images, could well represent the facial parts, the outlier pixels will
have rather big coding residuals. Thus , the pixel which has a big residual ei should have a small weight. Such a
principle can be observed from Eq.(11), where

(e
i
) is inversely proportional to e
i
and modulated by

(e
i
).
Refer to Eq.(11), since

is differentiable, symmetric, monotonic and has its minimum at origin, we can assume
that

(e
i
) is continuous and symmetric, while being inversely proportional to e
i
but bounded.without loss of
generality, we let

(e
i
)[0,1]. With these considerations, one good choice of

(e
i
) is the widely used logistic
function

(e
i
)=exp(-e
i
2
+)/(1+exp(-e
i
2
+)) (14)
Where and are positive scalars. Parameter controls the decreasing rate from 1 to 0, and controls the
location of demarcation point. Here the value of should be big enough to make

(0) close to 1.with

Eq.(14),Eq.(11),and

=0, we could get

(e
i
)= - (ln(1+exp(-e
i
2
+))-ln(1+exp)) (15)
The PDF f

associated with

in Eq.(15) is more flexible than the Gaussian and Laplacian functions to model
the residual e. It can have a longer tail to address the residuals yielded by outlier pixels such as corruptions and
occlusions, and hence the coding vector will be robust to the outliers in y.

(e
i
) could also be set as other
functions. However, as indicated by the proposed logistic weight function is the binary classifier derived via
MAP estimation, which is suitable to distinguish inliers and outliers.
When

(e
i
) is set as a constant such as

(e
i
)=2, it corresponds to the l
2
-norm fidelity in Eq.(3);when set as

(e
i
)=1/|e
i
|, it corresponds to the l1-norm fidelity in Eq.(4); when set as a Gaussian function

(e
i
)=exp(-
e
i
2
/2
2
), it corresponds to the Gaussian Kernel fidelity in FR. However, all these functions are not as robust as
Eq.(14) to outliers, one can see that the l
2
-norm fidelity treats all pixels assigns higher weights to pixels with
smaller residuals; however, the weight can be infinity when the residual approaches to zero, making the coding
unstable.
Both our proposed weight function and the weight function of the Gaussian fidelity used in FR are bounded in
[0,1], and they have an intersection point with weight value as 0.5. However, the proposed weight function
prefers to assign larger weights to inliers and smaller weights to outliers; that is, it has higher capability to
classify inliers and outliers.
There are also some candidates which could be adopted as the weight function of RRC. Like the Gaussian
weight function, these weight functions in M-estimation could also assign high weights to inliers and low
weights to outliers. Nevertheless, the proposed RRC model is a general model which could utilize various
weight functions, and in this paper we adopt the logistic weight function due to its advantage analyzed above.
The model in Eq.(4) is the case by letting

(e
i
)=1/|e
i
|.Compared with the models in Eqs.(3) and (4), the
proposed RRC model [Eq.(8) or Eq.(13)] is much more robust to outliers because it will adaptively assign small
weights to them. Although the model in Eq.(4) also assigns small weights to outliers, its weight function

(e
i
)=1/|ei| is not bounded ,making it less effective to distinguish between inliers and outliers.
Vol 05, Article 05363; May 2014

International Journal of Engineering Sciences Research-IJESR
http://ijesr.in/ ISSN: 2230-8504; e-ISSN-2230-8512


2010-2014 - IJESR
Indexing in Process - EMBASE, EmCARE, Electronics & Communication Abstracts, SCIRUS, SPARC, GOOGLE Database, EBSCO, NewJour, Worldcat,
DOAJ, and other major databases etc.,
1531

2.4 Two Important Cases of RRC
The minimization of RRC model in Eq.(13) can be accomplished iteratively, while in each iteration W and are
updated alternatively. By fixing the weight matrix W, the RRC with CGD prior on representation and o(j)=-
lnfo() could be written as
=arg min

{ ||W
1/2
(y-D)||
2
+ (|
j
|

+b
a0
)} (16)
where
o
(
j
)=|
j
|

+b
o
. =(1/
a
)

and b
o
=ln(2(1/)/) is a constant. Similar to the processing of
F

(e)=
i=1

(e
i
),
0
(
j
) could also be approximated by the Taylor expansion. Then Eq.(16) changes to
=argmin

{||W
1/2
(y-D)||
2
+ V
j,j

j
2
} (17)

where V is a diagonal matrix with V
j,j
=
0

(
j
)/
j
.
The value of determines the types of regularization. If 0 1, then sparse regularization is applied; otherwise,
non-sparse regularization is imposed on the representation coefficients. In particular, the proposed RRC model
has two important cases with two specific values of .
When =2, GGD degenerates to the Gaussian distribution, and the RRC model becomes
=arg min

{||W
1/2
(y-D)||
2
+||||
2
} (18)
In this case the RRC model is essentially an l2-regularized robust coding model.It can be easily derived that
when W is given,the solution to Eq.(18) is =(D
T
WD+I)
-1
D
T
Wy.
When =1, CGD degenerates to the Laplacian distribution, and the RRC model becomes
=argmin

{||W
1/2
(y-D||
2
+||||
1
} (19)

3. IR
3
C ALGORITHM
3.1 INPUT:
Normalized query image y with unit l
2
-norm; dictionary D;
(1).
3.2 OUTPUT:
Start from t =1:
1. Compute residual
e
(t)
=y D
(t).

2. Estimation weight as

(e
i
(t)
)=1/1+exp((e
i
(t)
)
2
)-)Where and could be estimated in each iteration.
3. Weighted regularized robust coding:
=arg min

{ ||(W
(t)
)
1/2
(y-D)||
2
+

(
j
)} (20)
Where W
(t)
is the estimated diagonal weight matrix with W
i,i
(t)
=

(e
i
(t)
).
0
(
j
)=|
j
|

+b
0
and =2 or 1.
4. Update the sparse coding coefficients:
If t=1,
(t)
=*;
If t>1,
(t)
=
(t-1)
+
(t)
(*-
(t-1)
);
Where 0<
(t)
1 is a suitable step size that makes

(y
i
-r
i

(t)
)+
0
(
j
(t)
)<

(y
i
-r
i

(t-1)
)+
0
(
j
(t-1)
).

(t)
can be searched from 1 to 0 by the standard line-search process .
5. Compute the reconstructed test sample: y
rec
(t)
=e
(t)
,And let t=t+1.
6. Go back to step 1 until the condition of convergence is met, or the maximal number of iterations is
reached.In this case the RRC model is essentiality the RSC model in FR, where the sparse coding
methods such as l
1
_l
s
is used to solve Eq.(19) when W is given.In this project,we solve Eq.(19) via
Eq.(17) by the iteratively re-weighting technique.

4. EXPERIMENTAL RESULTS
4.1 Pass the input image
Collect all the necessary images for the project and save it in the database
Use entropyfilt to create a texture image. The function entropyfilt returns an array where each output pixel
contains the entropy value of the 9-by-9 neighbourhood around the corresponding pixel in the input image I.
Entropy is a statistical measure of randomness.
Threshold the rescaled image Eim to segment the textures. A threshold value of 0.8 is selected because it is
roughly the intensity value of pixels along the boundary between the textures.


Vol 05, Article 05363; May 2014

International Journal of Engineering Sciences Research-IJESR
http://ijesr.in/ ISSN: 2230-8504; e-ISSN-2230-8512


2010-2014 - IJESR
Indexing in Process - EMBASE, EmCARE, Electronics & Communication Abstracts, SCIRUS, SPARC, GOOGLE Database, EBSCO, NewJour, Worldcat,
DOAJ, and other major databases etc.,
1532

Fig 4.1 Pass input image and feature extraction
4.2 Segment the image based on region



Fig.4.2 Segmentation of input image
Vol 05, Article 05363; May 2014

International Journal of Engineering Sciences Research-IJESR
http://ijesr.in/ ISSN: 2230-8504; e-ISSN-2230-8512


2010-2014 - IJESR
Indexing in Process - EMBASE, EmCARE, Electronics & Communication Abstracts, SCIRUS, SPARC, GOOGLE Database, EBSCO, NewJour, Worldcat,
DOAJ, and other major databases etc.,
1533
4.3 Find the intensity distribution

(a) (b)
Fig.4.3 (a) Gradient Magnitude (Gradmag) (b) Watershed transform Of Gradient Maagnitude (Lrgb)
Your first step is to maximize the intensity contrast in the image. You can do this using ADAPTHISTEQ, which
performs contrast-limited adaptive histogram equalization. Rescale the image intensity using IMADJUST so
that it fills the data type's entire dynamic range.
4.4 Segment the image based on region(with and without intensity)

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k)
Fig.4.4 (a)Opening Lo (b) Opening by reconstruction (Lobr) (c)Opening and closing of Loc (d) Opening and
closing reconstruction (lobrcbr) (e) Regional maxima of opening closing by reconstruction(fgm) (f) Regional
maxima superimposed on original image(L2) (g)Threshold opening-closing by reconstruction(Bw)
(h)Watershed ridge lines(bgm) (i) Markers and object boundaries superimposed on original image (j) Color
watershed label matrix(lrgb) (k) Lrgb super imposed transparently on original image.

Granulometry estimates the intensity surface area distribution of snowflakes as a function of size. Granulometry
likens image objects to stones whose sizes can be determined by sifting them through screens of increasing size
and collecting what remains after each pass. Image objects are sifted by opening the image with a structuring
element of increasing size and counting the remaining intensity surface area (summation of pixel values in the
image) after each opening. Choose a counter limit so that the intensity surface area goes to zero as you increase
the size of your structuring element. For display purposes, leave the first entry in the surface area array empty.

4.5 Comparison of Gaussian noise, speckle noise, and salt and pepper noise
Median filtering is a common image enhancement technique for removing salt and pepper noise. Because this
filtering is less sensitive than linear techniques to extreme changes in pixel values, it can remove salt and pepper
Vol 05, Article 05363; May 2014

International Journal of Engineering Sciences Research-IJESR
http://ijesr.in/ ISSN: 2230-8504; e-ISSN-2230-8512


2010-2014 - IJESR
Indexing in Process - EMBASE, EmCARE, Electronics & Communication Abstracts, SCIRUS, SPARC, GOOGLE Database, EBSCO, NewJour, Worldcat,
DOAJ, and other major databases etc.,
1534
noise without significantly reducing the sharpness of an image. In this topic, you use the Median Filter block to
remove salt and pepper noise from an intensity image

5. CONCLUSION
We have proposed a more general model for segmenting images with/without intensity inhomogeneities and
with different types of noise. Our proposed level set energy function, which is dominated by the global Gaussian
distribution and constrained by local neighbor properties, can overcome the artifacts from both the intensity
inhomogeneity and image noise. A quantitative comparison on synthetic images and experimental results on real
images showed that our model outperformed the LBF and LSII models designed specifically for segmenting the
images with intensity in homogeneities. It was more robust and accurate than the CV model when segmenting
the images without intensity in homogeneities.

6. REFERENCES
[1] Dexing Zhong,Peihong Zhu,Jiuqiang Han and Shengbin Li(2012)An improved Robust Sparse Coding
for Face Recognition with disguise,Int J Adv. Robotic Sy., Vol.9,126
[2] John Wright,A.Y.Yang,S.ShankarSastry,Arvind Ganesh and Yi Man(Feb2009),Robust face
recognition via Sparse representation,IEEE transaction Vol.31,No.2.
[3] M. Yang, L.Zhang,J.Yang, and D.Zhang(2011)Robust sparse coding for face recognition, Proc. IEEE
Conf. Computer Vision and Pattern Recognition.
[4] R. Chellapa, C.Wilson, and S. Sirohey. Human and machine recognition of faces: a survey.
Proceedings of the IEEE, 83(5):705741, 1995.
[5] C. Cortes and V. Vapnik. Support vector networks. Machine Learning, 20:125, 1995.
[6] M. Fleming and G. Cottrell. Categorization of faces using unsupervised feature extraction. In Proc.
IEEE IJCNN International Joint Conference on Neural Networks, pages 6570, 90.
[7] G. Guodong, S. Li, and C. Kapluk. Face recognition by support vector machines. In Proc. IEEE
International Conference on Automatic Face and Gesture Recognition, pages 196201, 2000.
[8] B. Heisele, T. Poggio, and M. Pontil. Face detection in still gray images. AI Memo 1687, Center for
Biological and Computational Learning, MIT, Cambridge, MA, 2000.
[9] K. Jonsson, J. Matas, J. Kittler, and Y. Li. Learning support vectors for face verification and
ecognition. In Proc. IEEE International Conference on Automatic Face and Gesture Recognition,
pages 208213, 2000.
[10] A. Lanitis, C. Taylor, and T. Cootes. Automatic interpretation and coding of face images using flexible
models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(7):743 756, 1997.



7. BIOGRAPHIES


V.Ezhilya received B.E degree in Electronics and Communication Engineering from
AnnaUniversity, Chennai in 2010 and M.E in Power Electronics and Drives from Vinayaga
Missions University. She is currently working as Head of the Department of Electronics and
Communication Engineering in VSA Group of Institutions, affiliated to Anna University.
Her major field of interest is Power Electronics, Digital Signal Processing, and Image
Processing. She was also a recipient of the distinguished graduate student award from
Vinayaga Missions University.