You are on page 1of 4

201O 2nd International Conference on Education Technology and Computer (ICETC)

Multi objective optimization for object recognition

Abdolhossein Alipoor
Science & Research Branch, Islamic Azad University
CE Department
Tehran, Iran
hosseinalipoor@srbiau.ac.ir

Mehdi Fesharaki

Maleke Ashtar University


CE Department
Tehran, Iran
fesharaki@mut.ac.ir

Abstract- Relevant with some important subjects like target


recognition, sensor fusion systems can be considered as one of

the main issues highlighting here. Environmental condition,

target characteristic and sensor efficiency are three parameters


which can impress on sensor value in target recognition so for
recognizing

targets, a group of sensors which have more

recognition rates, must be selected intelligently. Utilizing many


sensors to acquire the highest object recognition rate would

have extra cost and decrease energy of mobile sensors rapidly.

Therefore make a tradeoff between sensor numbers and object

recognition rate would be imperatively. This paper attempts to


design

multi

optimization

objective

algorithm

and

optimization
neural

service

network.

by

This

using

service

specifies highest recognition rate for each distinct sensor


numbers. We propose multi objective optimization algorithm
to help accessing the best sensory configuration for a definite

environment

regarding

to

the

environmental

conditions,

sensors performance, and object features. Our multi objective

optimization algorithm has two functions. Genetic algorithm is

used

to

perform

as

one

of

functions

to

specify

object

recognition rates of each sensor group. Neural network is used

to perform as fitness function of each genetic algorithm

chromosomes.
determinant.

Another

Highest

function

recognition

is

rate

sensor

and

numbers

lowest

sensor

numbers are two objects which multi objective optimization

algorithm wants to make a balance between them. We define

500 different scenarios for 6 different sensors in different


conditions. Object recognition rate of each sensor is collected.

These rates are used for neural networks training process. By

defining new scenario and run multi objective optimization


algorithm in this scenario, this algorithm makes a Pareto front
between sensor numbers and object recognition rate. Finally

this algorithm finds, by distinct numbers of sensor, which


sensors by which recognition ability must be used to reach the

highest recognition rate.

Sensory network nodes can be different kinds of sensors like


radar, infrared camera, laser and etc. In order to decrease the
processing time and also increase the awareness of
environment, we need an automated sensor management.
Automatic sensor management is performed in level 4 of the
JDL model named process management [1]. The process
management stage is an ongoing assessment of the other
fusion stages to ensure that the data acquisition and fusion is
being performed in a way that will give optimal results. This
could also improve results by adjusting the parameters in the
fusion process, establishing a target priority or selecting the
sensors to give improved recognition rates [1]. Three
processes are done in this level:
1. Situation measurement, which is in relation with
transmission of sensory platform to the best possible
point in order to receive the most accurate and the least
expensive information of environment.
2. Scheduling of sensor's tasks which is used for multi
functional sensors.
3. Sensors selection, which is done in multi sensor and
multi-target environments. The best sensors selection
task for object recognition is performed by this part.
The sensor selection operation determines the best sensor
configuration in the current situation. In this paper we focus
on the last process of automated sensor management to use
the best sensory configuration in any situation automatically.
For each platform to perform sensory data fusion, frrst,
sensors are selected considering environmental conditions,
target characteristics and sensor efficiency. After that
intelligent sensor selection performed, data fusion results
have maximum accuracy based on the group of selected
sensors; Because of not using inexact or noisy sensors
information.

Keywords- Automatic Sensor Management; Intelligent Sensor


Selection;

Objects

Recognition

Rate;

multi

optimization; Genetic Algorithm; Neural Network.

I.

INTRODUCTION

The architecture of object recognition is based on


networks of sensors and data fusion. Each of these networks
covers a part of environment and includes sensory nodes and
their fusion operations. Most of these operations must be
automated for accelerating their necessary decision making.
This accelerates observation, orientation, decision, and action
(OODA) cycle and also developing common operational
picture (COP) for each agent in environmental monitoring.
978-1-4244-6370-11$26.00 2010 IEEE

II.

objective

SENSOR SELECTION OPTIMIZATION SERVICE

Due to different efficiency of sensors and also many


effects of sensor numbers on sensor selection, multi
objective optimization for sensor selection is so important.
Figure 1 depicts the object recognition rate based on the
number of sensors.

V2-70

20i 0 2nd international Conference on Education Technology and Computer (ICETC)


o

High

0.95

0.99

Fig.2. (A)

Object
Recognition
Rate

Low

Low

Number of Sensors

FIG. I. OBJECT RECOGNITION RATE IN FOUR CONDITIONS

According to figure 1, at the first and second states,


efficiency of system is decreased because of choosing bad
sensors for recognizing targets. Why the second state has a
poor efficiency is sensor selection without considering the
environmental conditions, target characteristics, and sensor
efficiency. In the third and fourth states, using the optimized
sensors caused to increase system efficiency. In the third
state, because of using more sensors in recognizing task, the
information processing is time consuming and also has
redundant results. In the fourth state, only a group of sensors,
which have the most recognition rates, are used for sensor
fusion. In this paper, our goal is to present such service to
access the highest recognition rate by lowest sensors
automatically.
Many techniques have been developed for general
optimized solution searching for sensor selection. These
techniques range from applying constraints on the objective
function to streamline the optimization process [2] to
applying advanced artificial analysis techniques such as
genetic algorithms and simulated annealing algorithms [3].
Unique optimization techniques have been proposed such as
particle swarm optimization [4] and cutting and surrogate
constraint analysis [5]. The point of this section is not to
review each search solution or optimization technique but to
identify the variety of algorithms available. Optimization
techniques are well documented [6]. Genetic algorithm uses
a set of chromosomes to present possible solutions for
solving problems. Each chromosome contains substrings
called genes, expressing variants of the problem space. In
this paper, genetic algorithm is one of two multi objective
optimization function, used for fmding the highest
recognition rate of each sensors group. Figure 2(A)
illustrates a chromosome that is a set of sensors defmed as
one possible solution. Each chromosome has 6 genes and
shows the recognition rate of each sensor. In this figure,
Sensors 3, 5, 6 are used and had 0.95, 1, 0.99 recognition
rates respectively.
S6
SI
S2
S3
S4
S5

,
)

High

Fig.2. (B)
FIG.2. (A) CHROMOSOME STRUCTURE, FIG.2. (8) THE GENETIC ALGORITHM
PROCESS.

Figure 2(B) illustrates the whole process of sensor


selection using genetic algorithm. The contribution of this
paper is the use of neural network as an estimator to evaluate
the fitness value of each genetic algorithm chromosomes. In
all previous methods at first made a fitness function and then
use optimization approach to fmd the optimal solution so this
method is the first utilizing neural network within an
optimization algorithm for sensor selection in environmental
monitoring or target recognition [3, 6 and 7].
For implementing optimization algorithm and showing
their results use MATLAB Optimization Tools. In genetic
algorithm, stochastic uniform method is used for selection
function. This method lays out a line in which each parent
corresponds to a section of the line of length proportional to
its expectation. The algorithm moves along the line in steps
of equal size, one step for each parent. At each step, the
algorithm allocates a parent from the section it lands on. The
first step is a uniform random number less than the step size.
For scaling, function rank method is used. This method
scales the raw scores based on the rank of each individual,
rather than its score. The rank of an individual is its position
in the sorted scores. The rank of the fittest individual is 1, the
next fittest is 2, and so on. Rank fitness scaling removes the
effect of the spread of the raw scores. For crossover function
tow point function with 0.8 crossover fraction is used. This
function selects two random integers 'm' and 'n' between 1
and Number of variables. The algorithm selects genes
numbered less than or equal to m from the first parent,
selects genes numbered from m+1 to n from the second
parent, and selects genes numbered greater than n from the
first parent. The algorithm then concatenates these genes to
form a single chromosome. Constraint dependent function is
also used for mutation with 0.2 mutation fraction [lO, 11].

V2-71

2010 2nd International Conference on Education Technology and Computer (ICETC)

III.

EVALUATING FITNESS VALUE WITH NEURAL NETWORK

The most important part of optimization problems is to


evaluate fitness of any solution which showing to what
extend optimized the selected solution. Since there is no
assessment function for acquiring the best configuration of
sensors on the basis of problem space variation, we utilize a
Multilayer perceptron neural network (MLPNN) for fitness
value estimation. MLPNN has 6 inputs, includes sensors
recognition rates, which are the second six genes of each
chromosome and has 3 layers. The first layer has 50 neurons;
the second one has 20 neurons and the last one haslneuron.
The first and second layer take advantage of Tansig
activation function and the third layer use pure line activation
function. Error back propagation training algorithm with 0.1
learning rate and The LEVENBERG_ MARQUART
function is used for the training part as illustrated in formulas
1, 2, 3. In this formula J is a Jacobian matrix and H is a
Hessian matrix. g is gradient and the fourth formula shows
how the network parameters are trained. Besides, the
network training stop criteria is specified as the le-05 mean
square error. Supervised learning model is used for neural
networks training by using MATLAB toolbox [8, 9 and 10].

{H
g

JTJ
JTe

Xk+1

IV.

Xk

(1)
-

[JT J + JiI]-lJTe

(2)
(3)

SPECIFICATION OF NEURAL NETWORK INPUT AND


OUTPUT IN THE CURRENT PROBLEM

To estimate the fitness value of each chromosome


accurately, at first the neural networks should be trained with
a set of real data. In this problem, the neural network is
trained with 500 samples. When neural network performs the
accurate estimation on real data, it presents suitable response
per each input. Recognition task in this paper is done using
6 sensors in limited geographical area. Sensors which are
utilized in this paper are:
S1. A radar system, which is an X band pulse, searches
radar with range 100 km, pulse width O.I J.ls and 1
minute direction angular extraction accuracy.
S2. A radar system, which is a Y band pulse, searches
radar with range 80 km, pulse width O.2J.ls and 1 minute
direction angular extraction accuracy.
S3. IR early warning which has 30 km range and 1
minute direction angle extraction and height.
S4. TV early warning which has 30 km range and 1
minute direction angle extraction and height.
S5. LASER early warning which has 30 km range.
S6. Range only which has 50 km range.
In order to get real data, sensors in 500 different
scenarios with different conditions are run and get target
existence recognition accuracy rate of each sensor group. We
use them as real data in neural networks training stage and
after that, the neural networks will be able to estimate the
output of sensors in new scenario [5, 7 and 8]. Different
conditions in which sensors produce data are as follow:

Radio electronic warfare


Infrared electronic warfare
Night or day
Bad climate condition
Existence of light jamming resources in the air

As an example we assume that it is a good climate


condition and in a sunny day, a target from east and in the
direction of sunshine comes to a region with different
sensors including radar, LR, IR, and TV. The target is
equipped with electronic warfare with the capability of bad
effect on X band radars. This scenario is one of 500
scenarios on which recognition accuracy rate is evaluated.
V.

EXPERIMENTAL RESULT

In this paper, multi objective optimization algorithm is


used for sensor selection. Neural network is utilized for the
purpose of estimating fitness function of each genetic
algorithm chromosomes. By using 500 different scenarios,
the training task of neural networks is done. The Levenberg
Marquart function is used for neural network training.
Consequently, the neural network approach obtained the
efficiency with error accuracy of 1.85e-05. New scenario is
defmed and genetic algorithm is run to found which the best
sensors group which has highest recognition rate with lowest
time consuming in this situation. After that, multi objective
optimization uses the genetic algorithm as one of its function
and defmed sensor numbers determinant as another function.
Multi objective optimization must be able to solve the
confliction between two objects, highest recognition rate and
lowest sensor numbers.
Figure 3 (a) shows the result of multi objective
optimization algorithm. As depicted in the Pareto front, the
end part of figure 3(a), dedicated fl for object Recognition
rate function and 12 for sensor numbers. This algorithm
fmds, by distinct numbers of sensor, which sensors by which
recognition ability must be used to reach the highest
recognition rate. As instance, as depicted in index 2 on
pareto front part in figure 3(a) if the sensor numbers are 4,
the highest recognition rate is 0.857 and sensors 1, 2, 3, 5 are
used with 0.21, 0.339, 1, 1 recognition rate respectively.
Figure 3(A) shows the Pareto front which has made by multi
objective optimization algorithm. As illustrated in figure
3(A) and 3(B) the algorithm specify that by utilizing sensors
1 and 2 the highest object recognition rate for the state that
you want to use just 2 sensors, obtain and this rate is 0.221.
For 3 sensors state, sensors 1, 2, 5 made the highest
recognition rate. This rate is 0.46. For 4 sensor state as it said
before, sensors 1, 2, 3 and 5 made the highest recognition
rate. This rate is 0.857. For 5 sensor state, sensors 1, 2, 3, 4
and 5 made the highest recognition rate. This rate is 0.869.
Finally for 5 sensor state, using all sensors made the highest
recognition rate. This rate is o. 959.

V2-72

20i 0 2nd international Conforence on Education Technology and Computer (ICETC)


..

Optimization Tool

File

[4]

Eberhart, R., and Kennedy, 1., "A New Optimizer Using Particle
Swarm Theory," Sixth IEEE International Symposium on Micro
Machine and Human Science,Nagoya,Japan,pp. 39-43,1995.

[5 ]

Osorio, M., and Hernandez, E., "Cutting Analysis for MKP," Fifth
IEEE Mexican International Conference in Computer Science, paper
Id: IO. I I09IENC.2004.1342620,pp. 298-303,2004.

[6]

Fletcher, R., Practical Methods of Optimization, 2nd ed., John Wiley


& Sons publishing company,book,Chichester,United states,1987.

[7]

Mushini,R.,and Simon,D., "On Optimization of Sensor Selection for


Aircraft Gas Turbine Engines," IEEE 18th International Conference
on Systems Engineering,Las Vegas,NV,pp. 9-14,August 2005.

[8]

Martin T. Hagan and Howard B. Demuth, Neural Network Design,


PWS publishing company, book, china, 1996.

[9]

P. Boilot, E. L. Hines,M. A. Gongora and R. S. Folland , "Electronic


noses inter-comparison, data fusion and sensor selection in
discrimination of standard fruit solutions," Elsevier Sensors and
Actuators journal,Volume 88,Issue I ,Pages 80-88, I January 2003.

Help
Op .....

Solver.

r;

amultiobJ Mulbobjectiw opbmiution ullng Genetic A.


..

Problem
cpmulbobJl

FltntufunCbon:

Number ofVJnablu: 6
(omtrllnts:

A:.

llnurmequlhtlts:
llnur equlhtJts:

Atq:

Bounds:

lower.

b:

btq:
Upper.

L
C
fuii.

111l

Run solvulndVlf:wruults

Use random stitesfrom previous run

PIU!t

ScortdrvtrSlty

Stop

Cumnt ittration: 102

Plrttofront
ClurResults

Pardofront function wlues and decision variables


Indtx ....

f1

f2

xl

x2

1(3

1(4

x5

x6

1--,;tO 8'I--;t
-;;;
' --"7
';;r.':::,:t-+----;t'--""""
"' ;r---;;',C'"I
1-----:1-7.
,-' ",:0:
,,1--3,--:,,:7.
, 3-,7
'.J""
. :t-+""",l1;;t, ""'""-;:;
,,,;!---';;I,.
0.22

50.959

0
0
. 920
5
.5
7
0.998
0.7360.5]90,.42

AvtfJ9tPlftto sprUiC

[10] J.W. Gardner, P. Boilot, E.L. Hines, "Enhancing electronic nose


performance by sensor selection using a new integer-based genetic
algorithm approach," Elsevier Sensors and Actuators journal, Volume
106,Issue I,Pages 114-12I, 29 April 2005.

I-----.:===

[ I I] Henrik Petersson and Roger Klingvall, " Martin Sensor array


optimization using variable selection and a Scanning Light Pulse
Technique," Elsevier Sensors and Actuators journal, Volume 142,
Issue 2,Pages 435-445,5 November 2009.

D.841-

"

'.

Fig.3. (a)
Pafelofronl

55

45

"-

4
35

25

L--709-08--;O7--:O6-O5-074-703O'2

2,

Object RecogMlonRale

FIG.3. (s)
Fig.3. (A),(B) the result of multi objective optimization algorithm

VI.

CONCLUSION AND FUTURE WORK

We conclude that utilizing genetic algorithm with neural


networks is a good approach for intelligent sensor selection.
The role of neural networks is to estimate the fitness value of
each chromosome in the genetic algorithm. With the use of
genetic algorithm the most suitable configuration of sensor is
obtained. Using different kinds of neural network and
compare their results to found the highest recognition rate for
every specific sensor numbers can be considered as future
work.
REFERENCES
[ I]

Duncan Smith and Sameer Singh, " Approaches to MuItisensor Data


Fusion in Target Tracking: A Survey " IEEE transactions on
knowledge and data engineering, vol. 18, no. 12, pp. 1696 - 1710,
DECEMBER,2006.

[2]

Debouk, R., Lafortune, S., and Teneketzis, D., "On an Optimization


Problem in Sensor Selection for Failure Diagnosis," 38th IEEE
Conference on Decision and Control, Vol. 5, pp.4990 - 4995, paper
Id: 1O.l109/CDC. I999.833338,1999.

[3]

Worden, K., and Burrows, A., "Optimal Sensor Placement for Fault
Detection," Elsevier Engineering Structures journal, vol. 23, no 8,pp.
885-90 I,August,200l.

V2-73

You might also like