Download pattern recognition - CIS @ Temple University

PATTERN RECOGNITION
Jadd Jennings
Intro to AI
December 13, 2004
1
PATTERN RECOGNITION
My research involves a subject of AI known as pattern recognition. Pattern
recognition is simply an AI topic that deals with the categorization and depiction of data
(patterns). Also known as computer vision and a process that makes it possible to watch
televisions, and use digital cameras.
The goal of this report is to reasonably present information in a matter that gives
the general reader no less than an adequate understanding of pattern recognition.
In today’s society, pattern recognition is very important in maintaining a complex
security system. The retinal scan is a machine that relates to pattern recognition because
it processes the patterns of blood vessels on the back of human eyes. These patterns will
then be characterized and matched up with the appropriate patterns. Another example of
pattern recognition is in fingerprinting. Typical crime solving units rely on pattern
recognition systems when processing and locating fingerprints. All humans have a unique
fingerprint and this trait is used to solve many crimes. With pattern recognition
technology, authorities are able to process and categorize fingerprint patterns.
There is useful information available relating to pattern recognition. With my
acquired information I hope to express what pattern recognition is, the processes of
pattern recognition, and examples of pattern recognition. My paper will also include a
section on biometrics (the field relating to fingerprinting). I hope to gain a lot from this
research and if all goes well I will have a stronger foundation of AI intellect.
1
IMAGE PROCESSING
Demonstrating how computer systems process images expresses the dynamics of
pattern recognition. Images are represented by an array of digits with in a computer
system. Each element of this array is known as a pixel. When a computer system
perceives a colored image’s pixel, the pixel is compromised of three layers. One layer for
red, green, and blue because the color perception is based on the function C(R, G, B):
where any color in the rainbow is a mixture of red, green, and blue. This mixture of
colors is often known as the color cube. In an 8-bit color image the red, green, and blue
layers can have values that range from 0 to 255. Monochrome images can be just as
important as colored images. They are sometimes known as grey level images because an
8-bit monochrome image will have 256 different shades of grey.
Digitization is the method of converting an image into an array of numbers. This
is done by graphing the image on a square grid and sampling the brightness of each point
on the grid. Sampling the grid points is a process called spatial quantization. Spatial
quantization makes it possible to produce an exact copy of the original image. The
accuracy of the new image depends on how many samples were taken; the more samples
the more accuracy. Shannon’s sampling theory states that an image can be reconstructed
identically if the image is sampled at the Nyquist frequency. The Nyquist frequency is a
sample rate that is double the highest frequency in the original image.
Grey level quantization helps boost a computer’s processing time, and conserves
storage. The brightness values within an image can have an infinite range. However, grey
2
level quantization restricts these values to a finite range. Restricting the brightness values
conserve storage; integers in the range 0 to 2^n require only n-bits to be represented. We
would need only 4 bits per point for an image with brightness values in the range of 0 to
16 as oppose to 32 or 64 bits.
When assigning integers to the brightness values, one may think of the linear
quantization method. In this method; an image with the range 0 to 255 (or we could say 0
to N) would be divided into 256 (or we could say N+1) equal intervals. There is also the
non-linear quantization method. Obviously, this method divides the range into unequal
values. Some instances require that grey levels be represented by more integers than
others because they are frequently occurring. For an example: humans have a limited
perception of brightness. It is more difficult to see the difference in darker tones because
our eyes have a logarithmic response to brightness. Therefore, grey scale images are
usually quantized based on a logarithmic scale, meaning more bits are used for the darker
range of coloring.
There are operations that can be executed on images called point processes. These
operations are based on arithmetic. The point process: Cij = Aij + b (i = rows j =
columns) would set image C equal to image A with an increase of brightness. Point
processes can also include operations on two images such as an addition. The sum of two
images would be as follows: Cij = Aij + Bij. Point processing is valuable when a
graphical representation of an image is being used. These graphical representations can
be known as grey level histograms. The grey level histograms are graphs of the pixels in
3
an image at each grey level. The changes that point processes bring are expressed well in
the grey level histograms.
There is a case of grey level quantization in which only 2 levels of grey are used.
Images in this case are known as binary images. In binary images all points in the
background have the value 0 and all points on the object (foreground) have the value 1.
Binary images are important because they efficiently save storage and processing time.
When dealing with binary image data it is important to understand the threshold value.
The threshold value is used to convert a grey level image into a binary image. The
way to convert a grey level image is to determine a threshold value from the range of
grey level values. This threshold value will be a boundary; all values less than it will be a
0 and all values greater than it will be a 1.
4
CLASSIFICATION
Classification is an important process of pattern recognition; it deals with the
categorization of data. The classification process sets standards of data to a certain
measurement or condition. The goal is to get the data into the smallest groups possible.
For an example: think of three different types of living beings; the human, the dog, and
the bird. With classification these beings could be to set three groups. One group would
be all beings that have wings; the birds would fall into this category. Another group could
be all beings that walk on four legs. There are many animals that walk on four legs but in
this case it’s just the dogs. The last category would be all living beings that walk on two
legs and this of course would be the category for humans. This example of classification
is good enough for this data, but the process can be much more complicated.
If we wanted to perform classification on a bigger group of data we can examine
character recognition. This recognition categorizes the set of all capital letters. Feature
extractions would include: number of tee junctions, number of end points, vertical
symmetry, and horizontal symmetry. See1 figure 1 for a decision tree showing the
classification process of the first half of the alphabet. Notice that there is still more
classification needed because there are two errors. The letters ‘H’ and ‘I’ fall into the
same class and this is understandable because they are the same letters when you rotate
either 90 degrees. However, ‘G’ and ‘L’ fall into the same class as well, and they are not
1
This tree is based on a Mike James diagram but I updated it to make it more efficient.
5
similar. Therefore, another feature extraction would be necessary in order make this
classification 100% efficient.
Even with more extraction; this classification is based on a categorizing a whole
image. The pixels are of no concern and there is no use of digitization anyways. The
very complex forms of classification deal with the objects that make up the image and
their relationships with each other.
FIGURE 1
ABCDEFGHIJKLM
0
Vertical Symmetry
(0 = no 1 = yes)
1
AFGJLM
BCDEHIK
0
GJLF
0
0
JF
1
BCDEK
AM
1
GL
Horizontal Symmetry
(0 = no 1 = yes)
1
0
2
M
A
# of Tee Junctions
1
0
CDK
# of End Points
HI
BE
D
0
3
2
2
3
0
B
C
J
F
D
6
4
K
E
IDENTITY VERIFICATION
The pattern recognition process is often used in our everyday lives. Digital
cameras, televisions and of course our very own CPU’s rely on pattern recognition. But
in my opinion the most notable uses of pattern recognition lie in identity verification. The
use of identity verification with AI pushed the human race to another level of
intelligence, identity verification is very important in our society.
One form of identity verification is the fingerprint verification. This verification is
used for security access and criminal investigations. To verify; one fingerprint (claimant)
is matched up with another fingerprint (enrollee) already on the database. The features of
fingerprints that are matched up are the ridges (lines on fingerprints) and valleys (spaces
in between ridges). A typical fingerprint image will consist of 8 bit pixels and 250 pixels
per inch. The images must be enhanced before any matching takes place. One step of the
enhancement process is the binarization step. This step is notable because the threshold
value is not constant. The inconsistent threshold value is caused by errors in the
fingerprinting capturing2. After binarization and the rest of enhancement the matching
finally takes place. The minutia matching is a system of matching that relies on the
bifurcation (forks) and ending features. During feature extraction minutias are located
throughout the image. The minutia locations are connected, stored as a graph and finally
matched up with the enrollee fingerprints minutia.
2
The contrast of fingerprints will be differing if the finger was pressed with more weight in the center.
7
CONCLUSION
Now that I understand complex techniques of pattern recognition my knowledge
of artificial intelligence has broadened. I especially understand how it can be said that
“artificial intelligence (you may have said robots) is making computers act like humans”
(Pei Wang). I agree with this because I have learned ways in which humans use pattern
recognition. When we were infants our first recognition is our mother. Recognizing a
known face puts our brains through a rapid recognition process. A slower process takes
place when we don’t immediately recognize someone. The point is that humans are born
with a natural intelligence and computers are born with an artificial intelligence. This is
why pattern recognition can be called “a computers vision.” If computers operate so
much like humans than who can deny that we as humans are not in some complex form
computers? Computers have limits and I believe human intelligence has limits as well but
we have yet to reach and understand these limits.
8
WORKS CITED
Mike James. Pattern Recognition. London: Wiley-Interscience, 1988.
Anil Jain, Ruud Bolle, and Sharath Pankanti. Biometrics: Personal Identification in
Networked Society. Boston: Kluwer Academic Publishers, 1999.
Julian Ashbourn. Biometrics: Advanced Identity Verification The Complete Guide.
London: Springer- Verlag, 2000.
Michael W. Davidson. “Basic Properties of Digital Images.” Molecular Expressions
Microscopy Primer (2002):
<http://www.molecularexpressions.com/primer/digitalimaging/digitalimagebasics.html>.
9
10
11