Académique Documents
Professionnel Documents
Culture Documents
Abstract The latest advancements in Science and vector in the image regions are left skewed. But in some
Technology, have witnessed radical changes in the image the pixel intensities of the image regions may not
banking system. Today most of the banks adopt the net be left skewed. They may have long upper tail with a
banking facility and most of the users are also addicted to right skewed nature. To segment these types of images, it
this system. Accordingly, most of the transactions are now is needed to consider that the pixel intensities of the
online based and much emphasis is therefore needed to image regions follow a right skewed distribution. Hence,
ensure the security of authenticating a person and in this research article we develop and analyze image
validating the transaction. Many models are therefore segmentation method based on mixture of Pearsonian
proposed in the literature. Most of these are an Type III Distribution. The Pearsonian Type III
alphanumeric based password schemes or biometric Distribution is capable of portraying right skewed and
schemes or graphical password based schemes. Each of long upper tail distributions. Here, it is assumed that the
these models is proposed by underlying an advantage. pixel intensities of each image region follow Type III
The alphanumeric passwords are proposed, with the Pearsonian Distribution and the pixel intensities in the
assumption that generating the password is easy and the whole image is characterized by a finate mixture of
generated password is unique, the biometric password Pearsonian Type III Model.
schemes are proposed with the assumption that tampering
a biometric is next to impossible. Graphical passwords II. PEARSONIAN TYPE III MIXTURE
are proposed with an option to the user so that he can DISTRIBUTION
select an image of his choice and then select some points In any image analysis, the entire image is considered as a
which is called a click pattern, which is unique to every union of several image regions. In each image region the
user. image data is quantified by pixel intensities. The pixel
Keywords Graphical Password Authentication, intensity z = f ( x , y ) for a given point (pixel) (x, y) is a
Pearsonian Type III Mixture Model, Statistics, random variable, because of the fact that the brightness
Probabilistic model, MIR Flickr. measured at a point in the image is influenced by various
random factors like vision, lighting, moisture,
I. INTRODUCTION- PEARSONIAN TYPE III environmental conditions etc. To model the pixel
MIXTURE MODEL intensities of the animal and human image regions, it is
In the earlier research work, we developed and analyzed a assumed that the pixel intensities of the region follow a
model for segmentation based on the mixture Pearsonian Pearsonian Type III Distribution (PTIIID). The
Type I Distribution with K-means algorithm. These probability density function of the pixel intensity is
models are useful when the pixel intensities of the feature
qi ai
q a l ai ai
z
f i z | ai , qi q a
i i
e q z l s
i s
,ai zs , q1
e aqi ai l
(Equation-1)
a1
i i
amma Function
N
L p Z s , 1
S 1
N
K
That is L 1 f i Z S ,
S 1 i l
This implies
K
K
L log i f i Z S ,
S l i l
i,i=1,2,K) is the set of parameters
N K (q a ) z
qi ai
log L( ) log qi ai i i i e qi zs 1 s
e ai (qi ai 1)
s 1
i 1 ai
(Equation-2)
The beginning step of the Expectation Maximization
algorithm requires the initialization of two parameters
( ai qi ; i=1,2,k) and component weights
Fig.1: Frequency Curves of Pearsonian Type III ( i ; i 1,2,..., K ) from the observed values. The idea
Distribution
of EM algorithm is to iteratively calculate Maximum
likelihood estimates of unknown parameters
The entire animal and human image is a collection of
regions which are characterized by Pearsonian Type III (ai qi i; i 1,2,...,K )
Distribution. Here, it is assumed that the pixel intensities
of the whole image follow a K-component mixture of E-STEP :
Pearsonian Type III Distribution and its probability In the expectation (E) step, the expectation value
density function is of the form: of log L( ) with respect to the initial parameter vector
(0)
K
pz i f i z | ai , qi
i l
Q ; (0) E ( 0 ) log L( ) / z
Where, K is number of regions, 0 i 1 are weights
Given the initial parameters
(0)
, one can compute the
such that i 1 and is f i z | ai , qi as given in
density of pixel intensity z s as
equation (1) i is the weight associated with ith region K N
p( zs , (1) ) i fi ( zs , (1) ),L( ) p(zs , (1) )
(1)
in the whole image.
In general the pixel intensities in the image regions are i1 s1
t ( z log
K N
Q ; , ) log f i ( z s ,
(1) (1) (1) (l )
i s i )) can be solved by applying the standard solution method
i 1 s 1
This implies
qi ai
qi zs zs
( qi ai 1)
K N
(1) (qi , ai ) 1 i 0
K
i s
i i1 s1
(t ( z , ) l o g
e qi ,ai ai (qi , ai 1)
e 1 l o g i
i 1
ai
This implies
Summing both sides over all observations, we get
N
1 N
i
N
t (z
s 1
i s , (1) )
Therefore
i th
The updated equation of for (l+1) iteration is
N
1
t ( z ,
( l 1)
i i s
(1)
)
N s 1
1 N ( l ) f ( z , (1) )
N
s 1
K
i i s
i
(l ) (1)
f i ( z s , )
i 1
(Equation-6)
We have
Q ; (1) E (log L( (l ) )
Therefore i
Q ; (1) 0
implies
E (log L( (l ) ) 0
K N (l )
i s
i i 1 s 1
(t ( z , (1)
)(log f i ( z s , (1)
) log i )) 0
K N qi ai
qi zs zs
( qi ai 1)
(qi , ai ) 0
(1)
log i
(l )
(t ( z , )
e 1
i i1 s1 log e ai (qi , ai 1)
i s qi ,ai
ai
K N z
i
(q a 1) log(q a ) log a q a log((q a 1)) q z
i i i i i i i i i i s qi ai log1 s ti ( z s , (1) ) 0
i 1 s 1 ai
N
(qi ai 1) z zs
(q ) log(q a )
i i i qi qi qi log1 s qi (qi ai 1)ti ( zs , ) 0
ai ai zs
(1)
s 1 qi ai
N a zs ai z s
q qi (qi ai 1)ti ( z s , ) 0
1
(1)
logqi ai i
ai z s
i
s 1
ai ai
N
ai z s 1 N
s 1 ai z s
qi (qi ai 1) qi log(qi (ai z s )) ti ( z s , (1) ) t ( z ,
i s
(1)
)
ai s 1
N
t i ( z s , (1) )
ai
s 1 ai z s
qi ( qi ai 1) qi log( qi ( ai z s )) t i ( z s , (1) )
ai z s
The updated equation of ai at (l+l)th iteration is
ai l l
N
t i Z s , l
l
qi z s
s 1
l qil qil ail 1 q logq a a
l l l
z s ti z s , 1
ai z s
i i i i
i1 f i z s , (1)
Where ti z s , (1)
f z ,
K
1 (1)
i i s
i 1
z
qi ai log1 s ti zs , 1 0
K N
qi
(q a 1) log(q a ) log a q a log((q a 1)) q z
i i i i i i i i i i s
i 1 s 1 ai
zs
K N
qi
( q a
i i 1) log( q a
i i ) log ai q a
i i log( ( q a
i i 1)) q z
i s q a
i i log 1
1 ti zs , 0
i 1 s 1 ai
l
qi ai zs
z a e log z dz
log( q a ) (qi ai 1)ai q z a log 1 zs 0
N s i s s
t z , 1 0
i i i i s i l
i s
qi ai ai
0 zs aie dzs
s 1 qi ai zs
l
qi ai zs
z a e log z dz
log( q a ) (qi ai 1) (q z ) a log 1 zs 0
N s i s s
t z , 1 0
i i i i s i l
i s
qi ai ai
s 1 qi ai zs
zs ai e dzs
0
zs ai (qi ai 1) ai (qi ai 1) 0 (qi ai 1)
N
ai log qi ai (ai zs )
0 (qi ai 1)
1
ti zs , 0
s 1 ai qi
N z ai q a 1 a q a 1t z , 1 0
ai z s i i
ai log qi ai s i i i i s
s l ai qi
qi ai z s ai
N
qi ai q a
i i 1 ai z s ai log a
s 1 i
qi 1 (Equation-7)
N
ai ti z s , 1
s 1
The updated equation of qil (l+1)th iteration is
l l l l l z s ail
N
qi ia qi ai 1 ai
l
z s ai
l
log qi ai
a l
s 1 i
q l 1 1 (Equation-8)
t z ,
i N
l 1
ai i s
s 1
il 1 f i z s , (1)
Where ti z s , (1)
K
i 1
i
l 1
f i z s , (1)
data set as initial clusters. These data points
IV. K-MEANS CLUSTERTING ALGORITHM represent initial cluster centroids.
The K-means algorithm is one of the simplest clustering 2. Calculate Euclidean distance of each data point
technique for which the objective is to find the partition from each cluster center and assign the data points
of the data which minimizes the squared error or the sum to its nearest cluster center.
of squared distances between all points and their 3. Calculate new cluster center so that squared error
respective cluster centers (Rose H. Turi, (2001)). K- distance of each cluster should be minimum.
means algorithm uses an iterative procedure that 4. Repeat step II and III until clustering centers do
minimizes the sum of distances from each object to its not change.
cluster centroid, over all clusters. This evaluational 5. Stop the process.
procedure consists of the following steps. In the above algorithm, the cluster centers are only
1. Randomly choose K data points from the whole updated once all points have been allocated to their