Abstract

Human fingerprints are rich in details called minutiae, which can be used as
identification marks for fingerprint verification. The goal of this project is to
develop a complete system for fingerprint verification through extracting
and matching minutiae. To achieve good minutiae extraction in fingerprints
with varying quality, preprocessing in form of image enhancement and
binarization is first applied on fingerprints before they are evaluated. Many
methods have been combined to build a minutia extractor and a minutia
matcher. Minutia-marking with false minutiae removal methods are used in
the work. An alignment-based elastic matching algorithm has been
developed for minutia matching. This algorithm is capable of finding the
correspondences between input minutia pattern and the stored template
minutia pattern without resorting to exhaustive search. Performance of the
developed system is then evaluated on a database with fingerprints from
different people

1
1.1 INTRODUCTION

As we enter the new millennium, identification is used for

almost everything nowadays. With technology on a rapid pace of
improvement, the need for accurate identification is very
necessary, especially with the increased use of automated
transactions.
Humans have certain unique properties to them. The study of
these properties is called biometrics. Biometrics is the ability to
automatically recognize a person using distinguishing traits such as
fingerprints, face, retina or iris from the eye, voice, and hand
geometry. Each of these methods of recognition has advantages
and disadvantages. Attributes such as cost, size, reliability,
operating environment, speed and accuracy help determine the
suitability for different applications. Without examining each of
these biometric recognition methods in detail, a case can be made
that fingerprint recognition has the broadest applicability for most
systems and is the best place to begin a search for an appropriate
biometric.
With the online shopping getting more and more popular,
transactions on the internet has brought many security problems to
surface as people often give out their personal information and

2
credit card number for each of their purchases. Many times, a
person’s credit card number can be stolen and misused. If a
biometric recognition system is applied, it can help reducing the
risk of shopping online to an extent.
The use of fingerprint recognition has existed as a means of
identification for many years. Not only fingerprints are more
accessible, but also fingerprint recognition systems generally have
lower costs, faster speed, and more reliability compare to other
biometric recognition methods.
Each person has a different pattern of fingerprint, and these
patterns are made of ridges, which make loops and whirls that are
unique to each person.

Figure 1-1 Different types of prints: arch, loop, whirl

3
2
 Fingerprints are commonly classified as 5 different types:
whorl, left loop, right loop, arch, and tented arch. For most
recognition systems, difficulties arise in distinguishing between
fingerprints of the same type. Many recognition systems used
neural networks to find minutiaes, the ridge ends or splits, in
fingerprints to match a fingerprint. Since the number of minutiaes
varies among fingerprint, many people have tried to exploit this
fact to enhance fingerprint matching. However, the image quality is
very critical to minutiaes detection as we will discuss later. The
features extracted can also vary in length depends on the type. This
has resulted in some difficulties in devising a classification system.
Our system is implemented so that it yields a fixed length of
features, and many classification schemes can be applied. It’s more
immune to noises because of the characteristic of Gabor filters that
the features only preserve directions and frequency at a location .
Compare our system and the minutiae-based systems, the
center point of the fingerprint is important to find for reference in
obtaining features. The orientation also plays an important role for
its accuracy. The algorithm for center point determination should
be consistent and a rotation algorithm should be applied for better
results.

5
 Many of the problems for fingerprint recognition systems are
related to obtaining the fingerprint images. Plastic distortions,
scanning artifacts, scanning resolution, and uneven pressure are
examples of such issues. A system has to address these problems in
order to maintain reasonable and consistent results.
Fingerprint recognition is a type of image processing that
requires memory and computation. In order to have a reasonable
performance in speed, a recognition system has to take into
consideration of the hardware constraints as well.

6
1.2 METHOD

1.2.1 PRE-PROCESSING

A scanner was not accessible, but we obtained a fingerprint

database consisting of 5 different people with 60 images for
each. The fingerprint images were 256x256 pixels in grayscale.

1.2.1.1 CENTER POINT DETERMINATION

Initially, a reference point, or center point, must be

determined. The following steps in the process will all be
referenced around this point. The center point is found at the
point of maximum curvature of ridge lines.

i) apply Wiener filter to the image. This filter performs 2-D

adaptive noise-removal. It uses a pixel-wise adaptive
Wiener method based on statistics estimated from a local
neighborhood of each pixel. The size of these
neighborhoods are 5x5 pixels.
ii) determine the numerical gradient of the image in the x and
y directions.
iii) apply the Wiener filter on the x and y gradients to enhance
them further.
iv) divide input fingerprint image into blocks of 10x10 pixels.

7
v) compute the slope perpendicular to the local orientation of
each block using the following formula:

vi) for blocks with slopes ranging from 0 to π/2, trace a path
down until slope that is not ranging from 0 to π/2 is found
and mark it.
vii) compute the slope in the negative y direction and output an
x and y position which will be the center point of the
fingerprint for block that has the highest number of marks.

Figure 1- 2 Center point of fingerprint

8
1.2.2 MINUTIAE DETECTION

A minutiae is a pattern of the flow of the ridges or valleys. A

fingerprint can have number of minutiaes. The minutiaes can
be distinguished and classified into many different types. See
the diagram below:

Figure 3 Minutiae Types

Since each fingerprint has its unique sets of minutiaes, many

automatic fingerprint matching systems are based on
minutiae matching. The traditional way is to use neural
networks, which often requires lots of computing power. The
quality of images also becomes very critical to minutiae
detection because most programs are written to trace down
the ridge lines. For example, a short noise line along a ridge
can likely be mistaken for a bifurcation minutiae. Therefore,
we took noises into consideration, and we tried to have more

9
 flexibility on the image quality by just searching one
minutiae point near the center point of the fingerprint1.

We first create a binary image from the scanned image. The

threshold is set so that the noise on the image is not as
apparent. Since we only need to find one minutiae point
within a small area near center, we have used a different
approach in our minutiae detection. One difference is that we
only look for ‘bifurcations’ and ‘dots’. Before we start the
minutiae search, those two patterns are predefined. During
the search process, we use the predefined patterns to scan
through the image from 4 different directions. Once we find a
minutiae, the process is stopped and the location of the
minutiae is recorded along with the type. (bifurcation or dot)

1.2.3 IMAGE ORIENTATION

A major obstacle in fingerprint recognition is that the images

obtained are not usually perfectly aligned. Usually rotation
and displacement of some sort is evident. Attempting to
compare two fingerprints with different orientations will
affect the result significantly. To adjust for this, rotation of
the input image is performed.

10
To rotate an image both the angle of rotation and the origin
point are needed. To find the angle of rotation, 4 points are
needed: center point of image in database and input image,
and location of the same minutiae point for both images.
Then the angle is calculated using the following equation:

Angle of rotation = tan-1(y2-center_y2 / x2-center_x2)

– tan-1(y1-center_y1 / x1-center_x1);

The origin point to rotate around will be the center point of

the input image. The initial technique was to go through each
pixel of the input image, and use the rotation equations below
to find the corresponding rotated position (x’, y’). This
technique has problems though, when the corresponding
coordinates are found they are not usually integer values.
This implies that the resulting rotated image will only have
pixels at integer values, and that certain output pixels will not
get mapped by any input pixels. The resulting output would
be completely inaccurate.

x = R * Cos(a) y = R * Sin(a) x' = R *

Cos(a + q) y' = R * Sin(a + q)

11
 The method used to solve this problem is to do the reverse.
For every pixel in the rotated image, calculate which pixel in
the input maps to it. Doing this is simple because the
negative of the angle of rotation can be used. When mapping
back though the coordinates will also probably not be
integers, though the closest neighboring pixel can be used.
This implies that some output pixels will map to the same
input, but this is acceptable as it is very difficult to observe
and is significantly better than unmapped pixels.

2.1 MAIN PROCESS

2.1.1 SECTORIZATION AND NORMALIZATION

The fingerprint image needs to be normalized in order to that

no differences in brightness or contrast exist throughout the
image. These differences occur when acquiring the image
from a scanner or other digitizing device as well as the
pressure applied by the person.

Furthermore, sectorization is used in the next step in the

process –feature extraction. The 6 Gabor Filters used are
lined up with 12 sectors in each band.

12
The image is divided into 5 concentric bands, each with a
radius of 20 pixels, laid over the center point. The center
band has a radius of 12 pixels. The total diameter is
(20*5+12)*2-1 = 223 pixels. The five bands each contain 12
wedges. Therefore, there are 60 sectors for the image. The
center band is omitted, as its area is small relative to the other
bands. The radius values are specifically chosen so as to
avoid the effects of circular convolution.

2.1.2 GABOR FILTERING

After we finish normalization, we will then pass the

normalized image through a bank of 6 different angles of
Gabor filters. Each filter is created by producing a 33x33
filter image with angles 0, π/6, π/3, π/2, 2π/3 and 5π/6. By
applying Gabor filters, we are able to remove noises,

13
preserve ridge and furrow structures, and provide information
contained in certain direction of the image.

The Gabor filter has a height and width(33*33) to maintain

its peak center point. The parameters, δx and δy, the space
constants of the Gaussian envelope along x and y axis, were
set to be 4.0. The frequency is determined by the inverse of
the average inter ridge distance. Too large a value of f will
create spurious ridges, and close ridges will merge into one if
the value of f is too small. We found that the average distance
to be about 10 pixels.

14
 Figure 4 : gabor filter

We have considered applying FFT to the normalized image

and the filter, so that we can speed up the filtering by using
only multiplication. However, applying FFT takes extra
steps, and it also requires more memory space for coefficient
storage. As a result, we simply have the convolution done in
space. The sectorization is then used to detect the presence of
ridges in the direction of the corresponding Gabor filter.

2.1. 3 FEATURE VECTOR - VARIANCE CALCULATION

After obtaining the 6 Gabor filtered images, the variance of

pixel values in each sector is calculated. This variance gives
an idea of the concentration of ridges in each direction in that
sector. Higher variance values imply that the ridges in the
sector were in a more similar direction to that angle of the
Gabor filter, while lower variance values imply that the
ridges were in other directions and were thus filtered out. A
variance value was obtained for each sector and for 6 of the
filtered images. Therefore, a total of 60 variance values are
obtained and they represent the feature vector of the
fingerprint. The following equation is used to calculate the
variance for each sector:

15
 Fiθ - pixel values in ith sector for Gabor filter with angle θ
Piθ - mean of pixel values
Ki – number of pixels in ith sector

2.1.4 POST-PROCESSING
Mean Nearest Neighbor
We choose mean nearest neighbor as our classifier. One feature
vector from each individual person is calculated from the
training set and then saved in our database for future matching.
The feature vectors in the database are gathered from better
quality images. For a given fingerprint, we find that the
Euclidean distance between the target feature vector and feature
vector of each person in our database. The person that yields the
lowest distance is chosen and matched. As we have looked into
many classification schemes, the mean nearest neighbor requires
the least calculations to yield a result. It also consumes the least
storage space since it takes only one feature vector from each
person. Although this classifier does not take into account the
variations of fingerprints of the same person, it is sufficient
enough for our project.
2.1.5 ALGORITHM OVERVIEW

16
Figure 2- 5: ALGORITHM OVERVIEW
2.5 SIGNAL
FLOW

PC side EVM
“scanned” input Image rotation
Image and transformatio
n
Image centering
Sectorization and
Normalization
Minutiae locator.

Feature Vector Gabor Filters

Comparison

Figure 2-7 : flowchart of signal

17
18
3. Methods

Overall, the center point determination algorithm did not give

precise results. It was applied to the following 3 fingerprint images.
The corresponding center points are shown for each image
respectively. It is evident that the algorithm is not very accurate.
Noise and low image resolution are important factors that these
images possess. Better quality images may result in improved
center points. However, the algorithm used is not accurate in
general.

Figure 8 Results of center point algorithm

After performing a rotation on the image on the left, the resulting
image has a better orientation. However, since the input is 256x256
the quality of the rotated image is not as good as the input. Higher
resolutions would make the resulting image much smoother. Due to
the mapping of the pixels of such a low resolution image, the result
is not the same quality as the input. Nevertheless, the resulting
image is still acceptable and can be processed for fingerprint
matching. It would have been much better if higher quality images
were available.

Figure 9 Fingerprint before and after rotation

As for the minutiae detection algorithm, it did not function

very effectively. The main reason for this was the presence
of noise in the fingerprint images. The process takes in a
threshold value and binarizes the image. Therefore, noise
values are attributed to a 0 or 1 and affect the results greatly.
Minutiae detection is not a simple task and requires

20
extremely high quality images. If restrictions can be placed
on the quality of the input images then minutiae detection
could work. The

Figure 10 Results of minutiae detector

lighter stroked circle on the fingerprint image indicates a true

minutiae point, while the thicker stroked circle indicates a result of
the minutiae detector. This result is incorrect due to the presence of
noise in the image.

21
After performing, sectorization and normalization, the brightness
and contrast throughout the images became constant. There were
no variations apparent. The results of the 6 different Gabor filters
are shown in the figure below. By applying these filters, noise was
reduced, and the ridges in a similar direction to each filter were
preserved. A previous group performed Gabor filtering by using
the FFT and then the IFFT to obtain the result. By performing the
filtering in the time domain we were able to reduce the time and
memory requirements significantly.

Figure 11 Results of 6 Gabor Filters

22
Since we had a database of 5 individuals and 60 fingerprints each,
we partitioned these prints into a training set and a testing set. The
5 individuals all had the same type of fingerprint – loop. This made
it more challenging to do verification among the prints. The
training set consists of 15 prints that make up the database. They
are also used to determine the threshold for rejection. A few of the
remaining prints, many of which were of bad quality, were used as
the testing set.

In order to determine a threshold value for our verification process,

it was necessary to compare a fingerprint from one class to one in
each of the five classes. By looking at the Euclidean distance
values, it should be obvious to determine which class the
fingerprint belongs. Then after repeating the process a few more
times, a threshold can be estimated.

Many obstacles make fingerprint verification difficult such as

incorrect orientation, displacement, missing sections, and poor
quality. Orientation and displacement issues arise as differences
each time a fingerprint is digitized from the person. Perfect
alignment while scanning is not possible each time. When
verifying a fingerprint, it is rejected if its quality is poor causing
the classifier to give a Euclidean distance that is much greater than
threshold.

23
 Here is a sample of the results from the classifier when verifying a
fingerprint from Class 1:
Class Euclidean Distance
1 52195
2 95817
3 125299
4 86804
5 83989

As you can see the minimum distance among the results is for
Class 1. This implies that the fingerprint most closely matched
Class 1, which is correct. A thorough testing of the classifier was
not performed due to limitations in time. However, from the results
obtained the verification system seemed to perform extremely well
considering the fingerprints were all the same type.

This research project permitted us to learn topics discussed in class

as well as new topics. These include Gabor Filters, interpolation
methods (rotation), and classification techniques. For the future, an
improved center point determination algorithm that is more
accurate should be considered. The project turned out to be an
interesting experience in which team work and time management
were important.

24
MEMORY PAGING & PROFILE RESULTS

The following table shows the results of profiling the code

on the EVM:

Code No of cycles
Rotation 191887842
Sectorization/ Filtering/Feature 1.022 x 1010
Gabor extraction

Rotation – Internal Memory:

N/A

Rotation – External Memory:

Variable Size
Frame 256x256x4
Result 362x362x4
Points 8
Total 786328 bytes

Gabor – Internal Memory:

62x3x4 = 744 bytes

Gabor – External Memory:

Variable Size
25
Frame 255x255x4
Result 255x255x4
Gabor 33x33x4
Total 524,556 bytes

Due to the time limitation, not much optimization of the EVM code
was performed. Simple optimizations such as using a common
index for an array were done, but not DMA transfers. Further
optimization of both the rotation and the Sectorization/Gabor
Filtering code are definitely possible.

Conclusion and feature work

26
The reliability of any automatic fingerprint system strongly relies
on the precision obtained in the minutia extraction process. A
number of factors damage the correct location of minutia. Among
them, poor image quality is the one with most influence. The
proposed alignment-based elastic matching algorithm is capable of
finding the correspondences between minutiae without resorting to
exhaustive research. There is a scope of further improvement in
terms of efficiency and accuracy which can be achieved by
improving the hardware to capture the image or by improving the
image enhancement techniques. So that the input image to the
thinning stage could be made better, this could improve the future
stages and the final outcome

27
REFERENCES

1. Turn, turn, turn.

http://msdn.microsoft.com/library/periodic/period99/turn.h
tm
2. Fingerprint Recognition Through Circular Sampling.
http://www.cis.rit.edu/~dxc0331/web_thesis/thesis.html
3. Previous 551 project. Spring 1999 Group 19.
http://www.ece.cmu.edu/~ee551/Old_projects/projects/s99_19/
finalreport.html 4. Anil K. Jain, S. Prabhakar, Lin Hong,
"Fingercode: A Filterbank for Fingerprint Representation and
Matching," MSU CPS Technical Report Library.
http://www.cse.msu.edu/cgi-user/web/tech/document?NUM=98-
36
5. Fingerprint Verification & Recognition.
http://www.cs.earlham.edu/~odo/biometrics/

28