Genetic Algorithms For Object Recognition IN A: Complex Scene

GENETIC ALGORITHMS FOR OBJECT RECOGNITION
IN A COMPLEX SCENE
Daniel L. Swets, Bill Punch, and John Weng

A714 Wells Hall
Michigan State University
East Lansing, Michigan 48824
{ swet sd, punch, weng} Qcps. msu. edu
ABSTRACT
is laid out, a brief review of genetic algoritlhms is given.

Results are shown on this segmentation/recognition combination which demonstrate the utility o f t he approach.
A real-world computer vision module must deal with

a wide variety of environmental parameters. Object
recognition, one of the major tasks of this vision module, typically requires a preprocessing step t o locate
objects in the scenes that ought t o be recognized. Genetic algorithms are a search technique for dealing with
a very large search space, such as the one encountered
in image segmentation or object recognition. This work
describes a technique for using genetic algorithms t o
combine the image segmentation and object recognition steps for a complex scene. The results show that
this approach is a viable method for successfully combining the image segmentation and object recognition
steps for a computer vision module.
2. BRIEF REVIEW OF GENIETIC
ALGORITHMS
This section gives a brief overview of genetic algorithm
fundamentals. Goldberg [4] gives a wonderful introduction to genetic algorithms, and the reader is referred t o
this source for further information.
A genetic algorithm i:: an optimization technique
that operates on a populrvtzon of individual solutions.
Each individual solution, also called a strzng in the population, represents a proposed solution t o the problem
being solved. The theories of natural selection are applied to i,his population, and subsequent generations
of the population are obtained by applying selection,
reproduction, and mutation operators (among possible others) t o the population. With these operators,
the population of solutions is gently pushed towards a
good-and hopefully optimum-solution to the problem.
The CA designer provides a fitness function t o evaluate the fitness of each individual solution; this fitness
function is used t o propagate good indiividuals into
the next generation. Some set of these fit individuals are chosen for a crossover operation, which recombines the strings of the parents into new children solutions, trying t o build up healthier strings in the process.
The mututzon operator randomly alters some element
of an individual in order to further enhance the population, though typically only rarely in comparison t o
the crossover operation.
Crossover is the name given t o a simple reproduction operation because of the way that the parent strings
are recombined Usually one or two common points in
a pair of parents are chosen at random. For a one-point
crossover, the portion of the parent strings t o the right
1. INTRODUCTION
A central task of the computer vision module is t o recognize objects from images of the machines environment. Navigation systems require the localization and
recognition of landmarks or threats; robotic systems
must find objects t o manipulate; image retrieval systems require the localization and recognition of objects
in images in order t o find database records of interest.
Image segmentation is often a prerequisite to object
recognition [l]; various techniques have been proposed
which combine the segmentation step with the object
recognition itself [2] [3].
Genetic algorithms [4] have been used in computer
vision systems for such tasks as parameter tuning (e.g.,
[ 5 ] ) and feature extraction [6]; typically the actual image processing tasks have been handled by other standard methods.
This work describes using genetic algorithms for
object recognition, combining the image segmentation
task with the object recognition task to be solved in its
entirety by the genetic algorithm. Before the approach
595
0-8186-7310-9/95$4.00 0 1995 IEEE
of the crossover point can be called the crossover area;

for a two-point crossover, the area between the points
is the crossover area. One child is formed by taking the
crossover area from one parent and the non-crossover
area from the other parent-recombining the parent
strings. The other child is formed by reversing the process. The theory behind doing this operation is called
the schema theorem [4]; short, low-order portions of
strings contributing to fit individuals are thus created.
There are several issues of paramount importance
in the design of a genetic algorithm. First and foremost, the fitness function of the system must be designed appropriately to select good individuals, since
it is the only window that the population has to the
outside world. Furthermore, the representation of the
strings making up the population of individuals must
be selected with great care. The way that strings recombine and are propagated into the next generation
is inherently linked to the way they are represented
and interpreted. Finally, the selection of appropriate
genetic operators for the system is fundamental to obtaining a good solution in a reasonable time.
3. GENETIC ALGORITHMS FOR OBJECT
RECOGNITION
The task of locating and recognizing a particular object

of interest in a complex scene is quite simple when cast
in the framework of genetic algorithms. A brute-force
method for finding an object in a scene is to examine
all possible subimage positions and sizes. A subimage
extracted in this way could then be used as a database
query for the model database; the subimage/model pair
with the highest confidence is taken to be a good segmentation and recognition result. Because of the enormous computational complexity of this approach, however, intelligent researchers immediately dismiss this
approach as far too expensive. When cast in the framework of the genetic algorithm methodology, however,
this brute-force method can be elevated to the level of
an elegant solution t o the difficult image segmentation
and object recognition problem.
Since the genetic algorithm approach does well in
very large search spaces by working with only a sample
of the available population, the computation limitation
of the brute-force method using full search space enumeration does not apply. It is this brute-force method,
cast in the genetic algorithm methodology, that is explored in this work.
3.1. Genetic Algorithm Parameters
Let ( u z , u y )and ( & , b y ) represent the upper left and

lower right corner of the object of interest, respectively.
Then for an input image whose dimensions are h x w,

u x E [o,w],uy E [O,h],bz E ( u z , w ] , a n d by E (uy,h].
The encoding recommended by Punch et al. [7] provides a relative encoding of the coordinates on the
strings of the genetic algorithm population, and is used
for this work. The layout of each chromosome is given
in figure 1. The y-component of the lower-right corner is calculated using a standard aspect ratio for the
images being examined.
The evaluation function extracts the coordinates
from each chromosome in the population, performs the
subimage extraction, subimage resize, and probes the
image database for a likely match. The fitness of the
individual that defines the subimage is given as the
confidence that the database retrieval utility has for its
best match.
3.2. Database Retrieval Utility

The database retrieval utility described in [2] is used
for this work. This content-based image retrieval system is broken into two phases, a training phase and a
recognition phase.
In the training phase, a labeled set of images are
given to the system. Each of these images has been
previously pre-segmented such that only a single object of interest is contained in the image. Furthermore,
this object in the image has a standard size, position,
and orientation; the image dimensions for all of the
images are identical in this phase. The training of the
system consists of generating a hierarchical set of Most
Expresszve Features (MEFs) and Most Dzscrzminating
Features (MDFs) [2][8]for each object placed into the
model database. The theories of optimal linear projection are used in order t o fully automate the selection
of these features. The MEFs best describe a particular
object class; the MDFs optimally discriminate among
the various classes in a linear sense. These feature sets
are used to build a hierarchical space decomposition of
the image database.
In the test phase, an image probe is presented to
this hierarchical space decomposition. The Most Discriminating Features selected during the training phase
provide the basis for selecting which branch of the database hierarchy to explore at each level. When the
database hierarchy exploration reaches a leaf node, the
retrieval technique provides a confidence measure by
comparing the search probe and the database object
class defined by the leaf node in the space of the Most
Expressive Features.
This confidence measure returned from the database
retrieval algorithm for test probe 5 and database sam-
596
image center
[O,
Figure 1: Layout of chromosomes. k is the number of bits required to store the maximum of {UL,, iU&, BR,}
constraints whatsoever were placed on the images or
the objects being considered for the recognition.
ple p is given by
5. CONCLUSIONS AND FUTURE WORK
where z , p are the MDF or the MEF subspace vectors, and C is a diagonal matrix where each diagonal
element d, is the variance in the i t h dimension of the
subspace being used. 0 < p is a penalty function for
illegal point combinations and ensures that invalid solutions receive little weight in the population of possible
strings. We do not wish t o simply remove these invalid
solutions because they may contain building blocks useful for obtaining fitter individuals in subsequent generations. c(z, p ) always falls in the range of [O,2], and is
maximized when the test probe most closely matches
a database sample in the feature space.
This work proposes a technique for image segmentation

and object recognition using genetic algorithms on the
images directly. The experiments show ithat the genetic
algorithm performs well in finding areas of interest even
in a complex, real-world scene. Genetic Algorithms are
adaptive t o their environment, and as such this type of
a method is appealing t o the vision community who
must often work in a changing environment.
Clearly further experimentation could be performed
in order t o show the general applicability of the technique for a widely varying set of images and objects in
the images. Grey coding the fields would greatly improve the mutation operation by ensuring that flipping
a bit only changes a fields value by one. The way the
described system is implemented, a bit changed by the
mutation operator could cause a large change in the
value.
DeJong-style crowding [4] could be used t o find multiple objects of interest in an image. !since a natural
scene probably contains many objects, the segmentation and recognition is necessary t o make the technique
useful for real-world images.
Timing improvements could be ma,de by utilizing
the implicit parallelization of multiple independent generations evolving at the same time. Different areas of
the search space are explored in each of these populations; occasional swapping of individuals among populations could be done t,o improve diversity and gently
push all the populations t o a global matximum.
4. EXPERIMENTS
The experiments were performed using the Genetic Algorithm Optimized for Portability and Parallelism System (GAlOPPS) developed at the Michigan State University Intelligent Systems Laboratory [9].
Among the experiments performed was an unconstrained natural scene experiment. This experiment
allowed the genetic algorithm t o operate on a natural
scene t o try t o locate and recognize a face in a crowd.
The crowd image with the face located and recognized
by the genetic algorithm is shown in figure 2. The
genetic algorithm parameters used are shown in figure 3. As this experiment shows, the genetic algorithm
is capable of finding a valid segmentation in a complex,
natural scene.
Because the GAlOPPS [9] system does not re-evaluate individuals which have already been examined,
this genetic algorithm approach examined a total of less
than 750,000 subimages. For the brute force method
t o enumerate and examine the entire search space explored by the genetic algorithm, a total of four billion
( 2 3 ) subimages would need t o be checked. So, though
a large computational effort was expended, less than
0.02% of the possibilities were considered using the genetic algorithm sampling technique. Furthermore, no
6. REFERENCES
[l] J. .J. Weng, SHOSLIF: The hierarchical optimal sub-
space learning and inference framework, Tech. Rep.

CPS 94-15, Michigan State University, Department of
Computer Science, A714 Wells Hall, East Lansing,
Michigan 48824, March 1994.
597
Figure 2: Face segmented and recognized by the genetic algorithm during the test phase.
P,
P,
I
I
I
1
Scale
Pop. size
Max gens
0.60
0.09091
33 bits
1.5
1000
750
Figure 3: Parameters used for the natural scene experiment. P, is the probability of crossover; P, is the probability for
mutation; 1 is the length of the chromosome used; Scale is the scaling factor used for the ratio of the best : mean individual
in the population. The mutation rate was set so that on average, three bits will be flipped per individual in each generation.
The crossover was restricted to crossing between fields on the chromosome.
D. L. Swets and J. J. Weng, SHOSLIF-0: SHOSLIF
for object recognition (phase I), Tech. Rep. CPS 9464, Michigan State University, Department of Computer Science, A714 Wells Hall, East Lansing, Michigan
48824, December 1994.
J. Weng, N.Ahuja, and T. S. Huang, Learning recog-
discovery and image discrimination, in Proceedzngs, 5th Int. Conf. on Genetic Algorithms, (UrbanaChampaign), 1993.
[7] W. Punch, E. Goodman, R. Averill, M. Pei, L. ChiaShun, D. Ying, D. Redder, and V. Kureichik, Research applications using genetic algorithms, Tech.
nition and segmentation using the creceptron, in Proc
Rep CAD-94, Michigan State University Genetic Al-
Internatzonal Conference on Computer Vzszon, pp. 121128, May 1993. Berlin, Germany.
D. E. Goldberg, Genetzc Algorathms an Search, Optamazataon, and Machane Learnzng. Addison-Wesley,
1989.
B. Bhanu, S. Lee, and J. Ming, Self-optimizing image segmentation system using a genetic algorithm, in
Proceedzngs, 4th Int. Conf. Genetac Algonthms, (San
Diego), 1991.
W. A. Tackett, Genetic programming for feature
gorithm Group, East Lansing, Michigan 48824, 1994.

181 Y. Cui, D. L. Swets, and J. J. Weng, Learning-based
hand sign recognition using SHOSLIF-M, in Proc. Internatzonal Conference on Computer Vaszon, 1995.
[91 E. D. Goodman, GALOPPS-Genetic ALgorathm Optamazed for Portability and Parallelism System. Genetic
Algorithms Research and Applications Group, Michigan State University, East Lansing, Michigan 48824,
2.20 ed., 1994.
598

Genetic Algorithms For Object Recognition IN A: Complex Scene

Uploaded by

Copyright:

Available Formats

Genetic Algorithms For Object Recognition IN A: Complex Scene

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Genetic Algorithms For Object Recognition IN A: Complex Scene

Uploaded by

Copyright:

Available Formats

GENETIC ALGORITHMS FOR OBJECT RECOGNITION

Daniel L. Swets, Bill Punch, and John Weng

is laid out, a brief review of genetic algoritlhms is given.

A real-world computer vision module must deal with

2. BRIEF REVIEW OF GENIETIC

of the crossover point can be called the crossover area;

The task of locating and recognizing a particular object

Let ( u z , u y )and ( & , b y ) represent the upper left and

Then for an input image whose dimensions are h x w,

3.2. Database Retrieval Utility

5. CONCLUSIONS AND FUTURE WORK

This work proposes a technique for image segmentation

space learning and inference framework, Tech. Rep.

nition and segmentation using the creceptron, in Proc

Rep CAD-94, Michigan State University Genetic Al-

gorithm Group, East Lansing, Michigan 48824, 1994.

You might also like