Implementation of Feedback Based Face Classifier for Content Based Face Retrieval System

jumentousmanlyInternet και Εφαρμογές Web

21 Οκτ 2013 (πριν από 3 χρόνια και 10 μήνες)

90 εμφανίσεις

1



Implementation of Feedback Based Face
Classifier for Content Based Face Retrieval
System



Ashish Surywanshi

Computer Science and Engineering

Sangam University, Bhilwara, Rajasthan, India

Email
-
ashishsurywanshi@gmail.com




ABSTRACT
-
Content
-
based face
retrieval

attracted many researchers of various fields. There exist many systems for face
retrieval meanwhile. Retrieval of faces from face archive using Suitable features extracted from the content of face is
currently an active research area. The CBFR pr
oblem is identified because there is a need to retrieve the huge databases
having faces efficiently and effectively. For the purpose of content
-
based face retrieval (CBFR) an up
-
to
-
date comparison
of state
-
of
-
the
-
art, low
-
level color and texture feature ex
traction approach is discussed. In this paper we propose A New
Approach for CBFR to recognize face image. This Approach is applied to improve retrieval performance of the algorithm
to work with CBFR in face recognition algorithm. Our aim is to implement tr
aditional CBIR algorithm and improve the
performance of face recognition. After improve this algorithm we implement the CBFR method for face recognition.
More over it need to check authenticity we compare the performance of CBFR with a traditional algorith
m PCA. PCA is
a statically works for face recognition.


Keyword
s
-
CBFR, PCA, Relevance feedback.


I.
I
NTRODUCTION


A.
Introduction
-



Relevance feedback (RF) techniques have been widely used in content
-
based image retrieval (CBIR) to bridge
the semantic gap
and improve retrieval performance. However, most RF techniques use intra
-
query learning (short
-
term learning) to perform query tuning in a single retrieval session. Recently, inter
-
query learning (long
-
term
learning) has been employed to further improve re
trieval performance. Among these, retrieval pattern
-
based learning
is the most effective. Here, we briefly review several such representative techniques. Retrieval pattern
-
based inter
-
query learning techniques aim to establish the relationship between the
current and previous query sessions by
analyzing image retrieval patterns. If the two sessions have similar image retrieval patterns, these learning
techniques assume that the user must be searching semantically similar images. As a result, images with sim
ilar
retrieval patterns are always returned as retrieval results [2].The latent semantic analysis method on the term
-
by
-
document matrix to learn a generalized relationship between the current query and the search history. He [4] uses the
semantic space to
store retrieval patterns of all the query sessions and find semantically similar images. Han [4] uses
the memory learning technique to form a knowledge memory model to store the semantic information and learn
semantic relations. Applying the statistical co
rrelation on the retrieval log to analyze the relationship between the
current and past retrieval sessions. However, all these learning techniques require a matrix of size N×N to store
feedback information, where N is the number of images in the database.
The matrix may also be sparse if all the
queries fall into several semantic categories. Moreover, erroneous feedback may also lead to the storage of incorrect
semantic information and degrade retrieval performance. In our approach, we propose a novel retr
ieval pattern based
inter
-
query learning scheme for CBIR [1].

2



First, we apply SVM
-
based low
-
level classification and semantic correlation
-
based high
-
level classification in
intra
-
query learning to perform query tuning in a single retrieval session. Second, we build a semantic matrix in
inter
-
query learning to store
retrieval patterns of a certain number of randomly chosen query sessions. Third, we
extract semantic features of the database images based on the historical feedback gathered from the semantic matrix.
Fourth, we update semantic features of the query image
by reinforcing semantically relevant features and
suppressing semantically irrelevant features using semantic features of the relevant and irrelevant images selected by
the user during RF iterations. Fifth, we apply the semantic correlation
-
based technique

to measure the similarity
between updated semantic features of the query and semantic features of each database image to return top retrieved
images [8].


B.
C
ontent Base

I
mage

Retrieval

(CBIR)


An
image retrieval

system is a computer system for browsing, searching and retrieving images from a large
database of digital images. Most traditional and common methods of image retrieval utilize some method of adding
metadata such as captioning, keywords, or descriptions
to the images so that retrieval can be performed over the
annotation words. Manual image annotation is time
-
consuming, laborious and expensive. So that, large amount of
research done on automatic image annotation. Additionally, the increase in social web a
pplications and the semantic
web have inspired the development of several web
-
based image annotation tools. The first microcomputer
-
based
image retrieval system was developed at
MIT
, in the 1980s, b
y Banireddy Prasaad, Amar Gupta, Hoo
-
min Toong,
and Stuart Madnick.
Image search

is a specialized data search used to find images. To search for images, a user may
provide query terms such as keyword, image file/link, or click on some image, and the system

will return images
"similar" to the query. The similarity used for search criteria could be Meta tags, color distribution in images,
region/shape attributes, etc.




Image Meta search
-

search of images based on related metadata such as keywords, text, etc.




Content
-
based image retrieval (CBIR)


the application of computer vision to the image retrieval. CBIR
aims at avoiding the use of textual descriptions and instead retrieves images based on similarities in their
contents (textures, colors, shapes etc.) t
o a user
-
supplied query image or user
-
specified image features.


Content
-
based image retrieval

(
CBIR
), also known as
query by image content

(
QBIC
) and
content
-
based visual
information retrieval

(
CBVIR
) is the application of computer vision techniques to s
olve the image retrieval problem.
The problem of searching for digital images in large databases. Content based image retrieval is opposed to
concept
based approaches.
"Content
-
based" means that the search will analyze the actual contents of the image rath
er than
the metadata such as keywords, tags, and/or descriptions associated with the image.


The term 'content' in this context might refer to:

(i)

colors

(ii)

shapes

(iii)

Textures or any other information that can be derived from the image itself.


CBIR is
desirable because most web based image search engines rely purely on metadata and this produces
a lot of garbage in the results. Humans manually enter keywords for images in a large database can be inefficient,
expensive and may not capture every keyword t
hat describes the image. Thus a system that can filter images based
on their content would provide better indexing and return more accurate results.

There is a growing interest in CBIR because of the limitations inherent in metadata
-
based systems, as well
as the
large range of possible uses for efficient image retrieval. Textual information about images can be easily searched
using existing technology but requires humans to personally describe every image in the database. This is
impractical for very large
databases, or for images that are generated automatically, e.g. from surveillance cameras.
It is also possible to miss images that use different synonyms in their descriptions. Systems based on categorizing
images in semantic classes like "cat" as a subcla
ss of "animal" avoid this problem but still face the same scaling
issues. The term Content
-
Based Image Retrieval (CBIR) seems to have originated in 1992, when it was used by T.
Kato to describe experiments into automatic retrieval of images from a database
, based on the colors and shapes
present. Since then, the term has been used to describe the process of retrieving desired images from a large
collection on the basis of syntactical image features. The techniques, tools and algorithms that are used derive
from
fields such as statistics, pattern recognition, signal processing, and computer vision.

3


II.
P
ROBLEM IDENTIFICATION AND METHODOLOGY


A.

Problem Domain


There are various techniques are implemented to detect the face or for face recognition system but they are not
much effective to search faces over the large database system in sort time limit. To detect or recognize face from a
set of faces thus we requir
ed a new technique to implement. To search faces from large database system in sort time
more accurately and effectively.


B.
S
olution Domain


According to our problem domain we are going to implement CBFR technique. That is a new technique which
is based

on content of face and found target face more accurately and more efficiently from the old systems to
compare. The performance we compare it with PCA face recognition system.


C.

E
xpected Outcome


In this project our main goal to achieve the following
facts:

1.

Implement the traditional CBIR algorithm.

2.

found the performance of the CBIR algorithm

3.

modify the CBIR algorithm to work with face recognition

4.

To check the authenticity of algorithm for that purpose. Implements a traditional face recognition algorith
m
PCA.

5.

Compare the performance of all algorithms.




III.
P
ROPOSED SOLUTION

A.
Proposed Approach


In our method shows some relevant Faces to the user based on the query Face the user selects. We have
proposed a new approach based upon Content based Face re
trieval system in efficient manner to face recognition. In
Proposed system we have prepared a database consist a number of Faces. The user has to input test Face (query
Face) and selected object Faces. Low level Features like color and texture are then ext
racted for test (query) Face
and object Faces. Similarities distances are measured and calculating linear Coefficient of Correlation between tests
(query) and object Faces. [1]


Now, here we include first the Traditional CBIR algorithm to retrieve Face fro
m large data base. Second
the PCA algorithm and third the proposed CBFR algorithm According to this Approach is following in next steps.


Algorithm I

Traditional CBIR Approach


Second algorithm describes the Traditional (old) CBIR Approach and following
some steps:

1.

Pre
-
processing:

Input various object Images O
i

Create B
m

Block Matrix. Calculate Mean μ of Block Matrices. Concatenate
all Block Matrices obtain.

2.

Feature Extraction:

The four techniques are used and individually the pixel to pixel distances a
re calculated and averages are
found out for the resultant Images. Rows denote the number of object in the images. There is a threshold value
for which the matching image detected. The matching criteria for the matching of the images are taken at
80percent
. Which can could he desired level of matching.


3.

Similarity Computation

Calculate Quadratic Distance between two color histograms is

d
2

(Q, I) = (H
Q
-
H
I
)

t

* A (H
Q
-
H
I
)

4




4: Combining the Feature


The above said features are combined to match the image.

The above similarity is measured are taken then
the averages are taken out for the final output.


Algorithm II

Principal Component Analysis (PCA)
Approach




The database processing





We compute the average face






Then subtract it from the training faces






Now we build the matrix which is
N
2

by
M






The covariance matrix which is
N
2

by

N
2



C=X . X
T



Find eigen values of the covariance matrix




The matrix is very large




The computational
effort is very big



We are interested in at most
M
eigen values




We can reduce the dimension of the covariance matrix



S= X
T

. X



Find the M eigen values and eigenvectors




Eigenvectors of
C
and
S
are equivalent



Build trans
form matrix
W
from the eigenvectors of
S



Compute for each face its projection onto the face space

5







Compute the threshold






To recognize a face






Subtract the average face from it





Compute its projection onto the face space






Compute the distance

in the face space between the face and all known faces






Reconstruct the face from eigen faces






Compute the distance between the face and its reconstruction





Distinguish between




Algorithm III

(New CBFR Approach)


Step 1: Preprocessing

1.

Input
various object Faces O
i
.

2.

Create Bm1*1 Block Matrices.

3.

Calculate Mean μ of Block Matrices and Concatenate.


6


Step 2: Feature Extraction

1.

Convert Bm1*1 Block Matrices “f & g” RGB from space to HSV from space, where f and g represent the
average values of vecto
rs.

2.

Extract feature vector V
j

from HSV space. And combined all color and texture features.


Step 3: Similarity Computation

Calculate Euclidean Distance then get Euclidean (D)


(Where f and g represent the average values of feature vectors respectively.)

W
hile (T
i
= =O
i
) // where T
i

is the test Face & O
i

is the object Faces.

Repeat above procedure for n object Faces. Now we have

“N


object Face and its Euclidean distance Matrices.

Calculate correlation coefficient


Step 4: Relevance Feedback

Mark all faces
as Relevant Faces as well as Irrelevant Faces and Forms set R
r

and R
i
. (Where R
r

ε T
r

and R
i

ε T
r
) and T
r

is Training Set.

Step 5:

1.

The set R
r

is modified by including the query Face selected by user in it.

2.

Calculate feature vector of R
r

and R
i
. Feature
vectors of F
r

forms the positive set and feature vectors of
F
i

forms the negative set of data points for training the classifier.

3.

These sets are then given as input to classifier (IKSVM). It draws the hyper plane on the basis of them.

4.

The feature set of T
r

is calculated and then fed to classifier for classification so that the hyper plane
formed can separate data points in training set as positive or negative. Those data points which falls on
the positive side of the hyper plane forms a set F
Tr

and similarl
y the negative side forms another set F
Ti
.
Since a Face is a data point in a feature space we use the term data point and feature vector
interchangeably.

5.

This Process is iterated many times.

Step 6:

1.

Collect the set F
Tr

from the last iteration of training.

2.

Plot all the data points in F
Tr

in the same space.

3.

Now, say user wants g numbers of Faces from the database which are most relevant to the query Face,
then on the basis of histogram intersection based similarity measure calculate the distance of query
data

point with data points in the set F
Tr
.

4.

Sort in descending order the distances as two of the most relevant Faces will have histogram distance
close to 1 and the dissimilar ones will have distance close to 0 and then retrieve first P no. Faces from
that sor
ted list.


B.
S
ystem Architecture


According to our need we design our system in three modules


1.

training of model

2.

testing of our model

3.

finally compare the models


1.

Training of model:
In this module we add Faces in our database to perform training for all
three models

2.

Testing of model:
In this module we test systems are work correctly or not.

3.

Model evaluation:

in this module we make comparison performance of PCA, Traditional CBIR and New
CBFR.

7





Figure
: shows the architecture of our system



IV.
I
MPLEMENTATION


Hardware Interfaces

Recommended



2.0 GHz Processor required (Pentium 4 and above)



Minimum 2 GB RAM



25 GB hard disk space

Software

Server Side



Operating System (Windows 2000 and above)



Microsoft Visual Studio 2008 Frameworks 3.5



MS SQL Server 2000

User Defined Classes and there Responsibility

S. No.

Class Name

Responsibility

1

Login

This class is responsible for authenticate user

2

MainMenu

It is MDI form to call all the GUI when ever needed

3

TrainCBFR

It is GUI class
designed to train CBFR data model for face detection

4

TrainPCA

It is a GUI class containing all methods and code to train PCA

5

RecognizePCA

It use the trained data model to recognize the face which one
selected to test using PCA

6

RecognizeCBFR

It use the trained data model to recognize the face which one
8


selected to test using CBFR

7.

Compare

Compare performance of both algorithm based on memory and
consumed

8.

Face Converter

This class helps to convert face to matrix and matrix to face



V.
R
ESULTS ANALYSIS

A.
R
esults

Compare (CBFR, CBIR and PCA) Table in Memory (KB)













Compare (CBFR, CBIR and PCA) Bar Chart for Memory in (KB)


Figure
:
Compare algorithm bar chart for memory



0
5000
10000
15000
20000
25000
30000
35000
40000
45000
50000
1
2
3
5
10
CBFR(N
ew)
CBIR(Ol
d)

No. of
Images

Algorithm

CBFR (New)

CBIR
(OLD)

PCA

1

35201

41459

39764

2

36698

42754

40876

3

38404

43316

41364

5

40275

45990

39864

10

42345

46215

40347

9


Compare Algorithms Table (CBFR, CBIR and PCA) in Times (Ms)












Compare (CBFR, CBIR and PCA) Bar Chart for Time in (Ms)


Figure
:
Compare algorithm bar chart for time


VI.
C
ONCLUSION

As we discussed in above Result section we found that proposed algorithm perform better
than

traditional
CBIR algorithm and PCA algorithm. Thus, we conclude that content based face retrieval (CBFR) system selects the
most informative faces with respect to the query face by ranking the retrieved faces.

We conclude the following facts:


0
200
400
600
800
1000
1200
1
2
3
5
10
CBFR(Ne
w)
CBIR(Old
)

No. of
Images

Algorithm

CBFR
(New)

CBIR
(OLD)

PCA

1

11

70

321

2

12

77

525

3

15

78

859

5

55

243

878

10

89

345

1127

10



Algori
thm

Memory

Time

Accuracy

PCA

Medium

High

Low

Traditional
CBIR

High

Medium

medium

New CBFR

Low

Low

High

Our implemented model performs much well than traditional model of PCA and Traditional CBIR. It
consumes less memory than other implemented
model. In case of time taken our model consumes less than 20% of
time. And finally which factor is effect the whole things is that accuracy, because in this era of computer science
memory space is not an effective factor to compare with. We achieve 5
-
12% h
igher accuracy then the other two
traditional models. CBFR system is more robust and computationally efficient; complexity will be reduced in terms
of resources and decreasing semantic gap.

VII.
F
UTURE WORK

In future we work with similar algorithms and
elaborate more about CBFR with the below given proposed
work. In future we work on “Content
-
based Face Retrieval System Using Sketches” and Retrieving Faces from the
WWW (World Wide Web). Propose new Security policies for the Database and to store trillion

of Faces and retrieve
faces efficiently or effectively.



VIII.
R
EFERENCES



[1]. Neetsh Gupta, Dr. R. K. Singh. “A New Approach for CBIR Feedback based Image Classifier”. In The
International Journal of Computer Applications (0975


8887) Volume 14


No.4
, January 2011.

[2]. Shahrooz Nematipour, Jamshid Shanbehzadeh, Reza Askari Moghadam, “Relevance Feedback Optimization in
Content Based Image Retrieval Via Enhanced Radial Basis Function Network”, Proceedings of the International
MultiConference of Enginee
rs and Computer Scientists 2011 Vol I, IMECS 2011, March 16
-
18, 2011, Hong
Kong.

[3]. Prof. Sunita Chand, Mr. Aditya Varshney, “APPLICATION OF IMAGE RECOGNITION SYSTEM”,
JOURNAL OF COMPUTING, VOLUME 3, ISSUE 5, MAY 2011, ISSN 21519617.

[4]. Mohammod Abul K
ashem, Md. Nasim Akhter, Shamim Ahmed, and Md. Mahbub Alam “Image Recognition
System Based on Principal Component Analysis (PCA) with Back Propagation Neural Networks (BPNN)”,
Canadian Journal on Image Processing and Computer Vision Vol. 2, No. 4, April 20
11.

[5]. Shylaja S S*, K N Balasubramanya Murthy, S Natarajan, Nischith, Muthuraj R, Ajay S, “Feed Forward Neural
Network Based Eye Localization and Recognition Using Hough Transform”, (IJACSA) International Journal of
Advanced Computer Science and Applica
tions, Vol. 2, No.3, March 2011.

[6]. Prashant Sharma, Amil Aneja, Amit Kumar, Dr.Shishir Kumar, “Image Recognition using Neural Network and
Eigenvalues with Distinct Block Processing”, International Journal of Scientific & Engineering Research
Volume 2, I
ssue 5, May
-
2011 1 ISSN 2229
-
5518.

11


[7]. Hui Hui Wang, Dzulkifli Mohamad, N.A.Ismail. “Towards Semantic Based Image Retrieval: A Review”. In
The 2nd International Conference on Digital Image Processing (ICDIP 2010), Singapore, 2010.Proceedings of
SPIE, Vol
5511, 2010.

[8]. Neetesh Gupta, Niket Bhargava, Dr. Bhupendra Verma, Md.Ilyas Khan, Mr. Shiv Kumar “A New Approach For

CBIR Using Coefficient Of Correlation”, 2009 International Conference on Advances in Computing, Control,
and Telecommunication Technologi
es.

[9]. P.S.SUHASINI , Dr. K.SRI RAMA KRISHNA, Dr. I. V. MURALI KRISHNA “CBIR USING COLOR
HISTOGRAM PROCESSING”, Journal of Theoretical and Applied Information Technology 2005
-

2009
JATIT.

[10]. Rajshree S. Dubey Rajnish Choubey Joy Bhattacharjee, “Multi

Feature Content Based Image Retrieval”,
Rajshree S. Dubey et. al. / (IJCSE) International Journal on Computer Science and Engineering, Vol. 02, No.
06, 2010, 2145
-
2149.

[11]. C˘at˘alin
-
Daniel C˘aleanu and Corina Botoca, “C# Solutions for a Image Detection

and Recognition System”,
FACTA UNIVERSITATIS (NIˇS) SER.: ELEC. ENERG. vol. 20, no. 1, April 2007, 93
-
105

[12]. Nenad S. Kojić, Slobodan K. Čabarkapa, Goran J. Zajić, and Branimir D. Reljin, Senior Member, IEEE,
“Implementation of Neural Network in CBIR S
ystems with Relevance Feedback”, JOURNAL OF
AUTOMATIC CONTROL, UNIVERSITY OF BELGRADE, VOL 16:41
-
45, 2006.

[13]. Remco C. Veltkamp, Mirela Tanase “Content
-
Based Image Retrieval Systems: A Survey”, October 28, 2002
his is a revised and extended version of T
echnical Report UU
-
CS
-
2000
-
34, October 2000.

[14]. Katarzyna Agnieszka Olkiewicz, Urszula Markowska
-
Kaczmar, “Emotion
-
based Image Retrieval

an
Artificial Neural Network Approach”, Proceedings of the International Multiconference on Computer Science
and Inf
ormation Technology pp. 89

96, ISBN 978
-
83
-
60810
-
27
-
9 ISSN 1896
-
7094.

[15]. Samy Sade , Ayoub Al
-
Hamadi, Bernd Michaelis, and Usama Sayed “Image RETRIEVAL USING CUBIC
SPLINES NEURAL NETWORKS”, International Journal of Video & Image Processing and Network S
ecurity
IJVIPNS Vol: 9 No: 10

[16]. Dimo Dimov, Nadezhda Zlateva, and Alexander Marinov, “CBIR APPROACH TO IMAGE
RECOGNITION”.

[17]. Dr. H.B. KEKRE, Mr. DHIRENDRA MISHRA, Mr. ANIRUDH KARIWALA, “A SURVEY OF CBIR
TECHNIQUES AND SEMANTICS”, Dr. H.B. Kekre et
al. / International Journal of Engineering Science and
Technology (IJEST).