ALGORITHMS FOR BINARIZING, ALIGNING AND

RECOGNITION OF FINGERPRINTS

A. Pillai, S. Mil’shtein and M. Baier

Advanced Electronic Technology Center, ECE Dept.,University of Massachusetts, 1 University Ave, Lowell, MA U.S.A.

Keywords: Fingerprints recognition, Binarizing, Alignment, Fourier transform.

Abstract: Minutiae based algorithms are widely used today for fingerprint authentication. In this study, we report the

use of the Fast Fourier Transform (FFT) as a base principle for our recognition method, and have also

developed image normalization methods. We also developed a novel method to align fingerprints to a

common reference orientation based on the Fourier Mellin Transform. Two methods for image recognition

are described. The first method uses image subtraction techniques in conjunction with a thresholding

scheme. The second method, which is currently in development, utilizes multiple neural networks running

in parallel. This technique is expected to be able to run image comparisons on large databases in real-time

through the use of modern parallel processing technology. In this study we analyzed 720 fingerprints

generated by wet-ink, flat digital scanners, and by a novel touch less fingerprinting scanner. For the image

subtraction method comparing high quality fingerprints (prints taken in touch less way), the rate of success

is 97%. For poorer quality prints, (those taken with wet-ink) the rate of success dropped to 93%.

Recognition statistics are not currently available for the neural network based image recognition method as

it is currently in development.

1 INTRODUCTION

Fingerprints offer a unique method for personal

identification. Fingerprints afford an infallible

means of personal identification, because the ridge

arrangement on every finger of every human being

is unique and does not alter with growth or age.

Fingerprint authentication is the most preferred

method because of their distinctiveness and

persistence over time as specified by Maltoni

(2003). The individuality of fingerprints has been

discussed in detail by Pankati (2001). It has served

almost all the governments worldwide over many

years to provide accurate identification of criminals.

No two fingerprints have been found to be the same

in the billions of comparisons that have been done to

date unless they belong to the same person. It

outperforms DNA and other human identification

systems to identify more number of criminals.

The minutiae algorithm is widely used for

fingerprint authentication. Minutiae points are local

ridge characteristics that appear as either a ridge

ending or a ridge bifurcation. A complete fingerprint

consists of about 100 minutiae points in average.

The measured fingerprint-area consists in average of

about 30-60 minutiae points depending on the finger

and on the sensor area. These minutiae points are

represented by a cloud of dots in a coordinate

system. They are stored together with the angle of

the tangent of a local minutiae point in a fingerprint-

code or directly in a reference template. A template

can consist of more than one fingerprint-code to

expand the amount of information and to expand the

enrolled fingerprint area. In general this leads to a

higher template quality and therefore to a higher

similarity value of the template and the sample. To

overcome the drawbacks of minutiae, hybrid

methods have been proposed in Jain (2001).

There are many challenges that need to be

overcome when developing an algorithm which is to

be used for reliable recognition of fingerprints

recorded by different technologies. This is because

different fingerprint capture techniques create

different representations of a given finger. These

challenges include different format size of images,

non-linear distortions of fingerprint ridges,

differences in orientation, and variation of gray scale

values.

In the current study, we use the Fast Fourier

Transform (FFT) as a base principle for our novel

426

Pillai A., Mil’shtein S. and Baier M..

ALGORITHMS FOR BINARIZING, ALIGNING AND RECOGNITION OF FINGERPRINTS.

DOI: 10.5220/0003362104260432

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2011), pages 426-432

ISBN: 978-989-8425-47-8

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

recognition method. We have also developed a new

binarization method that is used to eliminate

variations in gray scale levels of each image, leaving

the resulting images looking like a traditional ink

rolled fingerprint. In this study we analyze 720

fingerprints generated by wet-ink, flat digital

scanners, taken from FVC 2004 and by the novel

contactless fingerprinting scanner described in

Palma (2006) and Mil’shtein (2008). In section 2,

we describe the binarization steps, section 3 contains

description about the fingerprint alignment process

and section 4 contains information about the

recognition procedure.

2 BINARIZATION PROCEDURE

Most fingerprint recognition algorithms currently

being used, including the widely used minutiae

algorithm, rely on the specificity of ridge endings

and ridge bifurcations. Because of this, it is

necessary to clearly define the fingerprint ridges and

valleys using only two distinct values, this process is

called binarization, and is one of the most important

steps that precede the recognition stage. Regardless

of the quality of any image recognition algorithm, a

poorly binarized image can compromise its

recognition statistics.

A good binarization algorithm would give an

image which would have very clear and uniform

black ridges on a white background even if the

image is overexposed to a certain degree. In the

current study the following binarization techniques

are used:

1) Region-Based thresholding

2) Filter-Based mentioned in Meenen (2005)

The first step is to divide the image into an N by

N grid of smaller blocks. Then the ridges like

regions within these smaller blocks are determined.

This is done by taking the gradients in the X and Y

direction and then finding the co-variance data for

the image gradients. Once this step is completed, the

orientation of ridges is computed. Estimation of

ridge frequencies in these blocks follows. For this,

first the mean orientation within the block is

obtained. Then the image block is rotated so that the

ridges are vertical. The rotated image is then

cropped so that it does not contain any invalid

regions. A projection of the grey values, down the

ridges, is obtained by summing down the columns.

Peaks in projected grey values are found by

performing grayscale dilation and then finding

where the dilation equals the original values. The

spatial frequency of the ridges is determined by

dividing the distance between the 1st and last peaks

by the number of peaks. If no peaks are detected, or

the wavelength is outside the allowed bounds, the

frequency image is set to 0. The ridges are then

enhanced with the help of a median filter. The image

obtained after this process is thresholded to obtain

the binary fingerprint. The threshold for binarization

depends on the resolution for the image.

We found that this technique works best with the

images that are obtained from the contactless

fingerprinting system described in Mil’shtein

(2008). This binarization technique is largely

invariant to inconsistencies in brightness levels

throughout the image, and results in a binary image

that has consistent information throughout. It should

be mentioned at this point that the downside of this

process is that a relatively large number of

calculations are needed, which adds to the time

needed for the overall recognition algorithm to

complete. In section 6 we give recommendations on

image registration procedures, but for now it is

recommend that all images that are to be stored in a

database be stored (at minimum) in their binary

forms to reduce computations when comparing

fingerprint images.

3 FINGERPRINT ALIGNMENT

Fingerprint alignment is an important stage that

precedes fingerprint recognition. It is important

because no matter which algorithm is being used for

recognition, one must be sure that the regions being

compared are the same. Fingerprint alignment using

eight special types of ridges extracted from thinned

fingerprint image is reported in Hu (2008) . Other

alignment techniques based on phase correlation of

minutiae points as described in Chen (2007), using

line segments as pivots based on minutiae as

mentioned in Carvalho (2004) and using similarity

histogram detailed by Zhang (2003), have also been

reported. But their inherent dependence on minutiae

necessitates a need for a new novel alignment

technique not based on minutiae. In this study, an

alignment technique based on the Fourier Mellin

Transform will be described.

The Fourier-Mellin transform is a useful

mathematical tool in image processing because its

resulting spectrum is invariant in rotation, translation

and scale. The Fourier Transform itself (FT) is

translation invariant. By converting the resulting FT

to log-polar coordinates, we can convert the scale

and rotation differences to vertical and horizontal

ALGORITHMS FOR BINARIZING, ALIGNING AND RECOGNITION OF FINGERPRINTS

427

Figure 1: Grayscale image of a selected fingerprint (Left) and the corresponding binarized image (Right).

Figure 2: Strength of binarization even if the image is seemingly overexposed.

offsets that can be quantified. A second transform,

called the Mellin transform (MT), gives a transform-

space image that is invariant to translation, rotation

and scale. An application of the Fourier-Mellin

Transform for image registration can be found in

Guo (2005).

The Mellin transform can be expressed as:

(1)

Convert to polar coordinates using:

(2)

We now have:

(3)

Making r = e

ỳ

and dr = e

ỳ

ỳ we have :

(4)

By changing coordinate systems from Cartesian

to a Log-Polar system, we can directly perform a

DFT over the image to obtain the scale and rotation

invariant representation. The figures below show

some of the results of the alignment using the

Fourier-Mellin transform on images taken from

Palma (2006).

The inverse Fourier transform of the Mellin

Transformed images helps to see how well the

image is aligned with respect to the base image.

While this step is necessary to see the alignment

results, the Fourier transforms are stored in as

separate database as from here they are now the

VISAPP 2011 - International Conference on Computer Vision Theory and Applications

428

Figure 3: Image of the 1

fingerprint.

Figure 4: Image of 2

fingerprint (In need of alignment).

templates that will be used for comparison. This will

eliminate the need to take again the FFT of the

aligned image and the base image when it comes to

comparing the fingerprints.

4 REGISTRATION

PROCEDURES

To run a fingerprint recognition system in its most

efficient state, steps should be taken within image

registration procedures to ensure that all possible

normalization of images and image transforms are

ALGORITHMS FOR BINARIZING, ALIGNING AND RECOGNITION OF FINGERPRINTS

429

Figure 5: Aligned 1

and 2

images (Image 2 superimposed upon image 1).

done upon image registration and not at the time of

image recognition. This way when it is time for

actual image comparison, fewer calculations will

need to be preformed, greatly increasing the

throughput of the algorithm. For this reason, we

recommend the following image registration

procedures.

It should be the normalized frequency domain

image that should be stored in the database rather

than the original or binarized image. This allows the

pattern recognition engine to directly take the image

data for use in comparison without the need for any

preprocessing steps. We recommend that as a new

image is acquired, it is processed and stored as

shown in figure 6.

Figure 6: Image registration procedure.

5 RECOGNITION ALGORITHM

There are three types of algorithms used for

fingerprint matching briefly described by Prabhakar

of MSU:-

a) Correlation based Algorithm:-

Here two fingerprints are superimposed on each

other and the correlation at intensity level between

corresponding pixels is computed.

b) Minutiae based Algorithm:-

Minutiae points are first determined on a fingerprint.

In order to make a comparison, 21 points are needed.

Minutiae points are nothing but points where there is

a ridge ending or a ridge bifurcation. In this process,

the minutiae points are stored as sets of points in a

two dimensional template. Then the algorithm finds

the alignment between the template and the input set

of minutiae sets that result in maximum number of

pairings. This stage requires the operator

intervention.

c) Ridge Feature based Matching:-

Here fingerprints are compared based on the features

extracted from the fingerprints.

In this study we report on the development of an

algorithm based on taking the Fast Fourier

Transform (FFT) of images and using a thresholding

scheme for comparison. It should be mentioned here

VISAPP 2011 - International Conference on Computer Vision Theory and Applications

430

that there are drawbacks to using a thresholding

scheme, and we are currently developing a neural

network based pattern recognition engine which can

take advantage of modern day parallel processing.

We expect to largely decrease the time necessary to

compare multiple images within a database,

hopefully opening up a door for large database, real-

time fingerprint recognition applications.

The Fourier Transform is an important image

processing tool which is used to decompose an

image into its sine and cosine components. The

output of the transformation represents the image in

the Fourier or frequency domain. The DFT is the

sampled Fourier Transform and therefore does not

contain all frequencies forming an image, but only a

set of samples which is large enough to fully

describe the spatial domain image. The number of

frequencies corresponds to the number of pixels in

the spatial domain image, i.e. the image in the

spatial and Fourier domain is of the same size.

For a square image of size N×N, the two-

dimensional DFT is given by:

(5)

Where f(a,b) is the image in the spatial domain

and the exponential term is the basis function

corresponding to each point F(k,l) in the Fourier

space. The equation can be interpreted as: the value

of each point F(k,l) is obtained by multiplying the

spatial image with the corresponding base function

and summing the result.

It is this frequency domain image that should be

saved as the reference information for registration of

a new fingerprint to a database. By saving this image

rather than the raw, oriented, or binarized images

which resulted from the previous steps one can

minimize the number of calculations necessary when

the time comes for actual fingerprint comparison.

Image comparison is currently done by comparing

the frequency domain images element by element to

see which have a similar value. Based on the

similarity of these values, a counter is incremented

and once this value crosses a statistically determine

threshold, a match is declared. We are working on

increasing the database so that we can provide ROC

and CMC datasets and curves and test the

performance on a wider database. For high quality

fingerprints (prints taken touch less way), the rate of

success is 97%. For poorer quality prints, (those

taken with wet-ink) the rate of success of our

algorithm is 93%. The advantage of this method of

comparison is that it is independent of minutiae

points and more reliant on the ridge patterns. If the

fingerprint is distorted with pressure induced

distortion then this algorithm can be used to

determine a match. We continue to increase the

database for analysis by our recognition algorithms.

A pattern recognition engine should be present in the

recognition stage to successfully compare an image

with a large database. The drawbacks of this

algorithm would be the time taken to come to a

decision and the requirement that the images

compared be of the same size.

We are currently developing a neural network

based approach for pattern recognition within

fingerprint images. Neural networks consist of

highly interconnected processing elements called

neurons which are all interconnected acting in

parallel to solve a common goal. This methodology

lends itself to parallel processing, and can take

advantage of modern day parallel processing

hardware, creating the opportunity for image

comparison in real-time even when comparing large

databases. The system in development currently

divides the images in a database into sets of

approximately 200 images, and trains one neural

network per image set utilizing an automated

training process described in Masters (1993). The

utilization of neural networks allows the system to

be optimized to run on high throughput graphics

processing units (GPU’s). By selecting the number

of neurons in the network to be equal to the number

of pixels in the frequency domain images being

compared, a direct image-to-image comparison can

be done in one GPU clock cycle. Each network is

capable of comparing approximately 200 images

accurately, and the number of networks in parallel is

limited by the number of GPU’s in the system. By

creating many neural networks, all autonomously

trained, we expect to be able to create a real-time

fingerprint recognition system for large-scale

databases.

In Gour (2010), the fingerprints are recognized

by using Monolithic and Modular Neural Network.

A monolithic neural network is one that takes an

input vector X and produces an output vector Y. The

relationship between X and Y is determined by the

network architecture. It generally consists of three

layers: one input layer, one output layer and more

than one hidden layer. The Backpropagation neural

network is an example of the monolithic network.

The training of a backpropagation network involves

three stages: the feedforward of the input training

patterns, the calculation and backpropagation of the

associated error and the weight adjustments.

Backpropagatin network is trained for query

ALGORITHMS FOR BINARIZING, ALIGNING AND RECOGNITION OF FINGERPRINTS

431

fingerprint to match it among a large number of

stored fingerprints as is evident in Jin (2002). The

average training time is 44.7 secs and the accuracy is

98%.

In the modular approach, one task is decomposed

into subtasks, and the complete solution requires the

contribution of all modules. To train a modular

neural network, which is having N number of

modules (feature points) in a particular fingerprint

requires two steps: Training of small modules and

training of intermediary modules. All the modules

are trained by using the backpropagation neural

network algorithm specified by Gour (2005). The

average time taken is 1.84 secs and the accuracy is

100%. Due to modularity, the modular neural

network gives better performance as compared to

monolithic networks.

6 CONCLUSIONS

We reported the development of a novel fingerprint

normalization and authentication algorithm which

has binarization, alignment, and recognition stages.

It is important to note that our method of fingerprint

image processing requires organization of database.

Structuring of database is orientation of all fingers

with regards to the position of the reference delta.

Although, we are suggesting a quality control in our

flow of processing to be done by Inverse Mellin

Transform, this step is more precautionary method.

Unlike, widely distributed minutiae based

fingerprint processing; our method does not require

interference of operator or final analysis by an

operator. We also continue to increase the database

so that we can provide ROC and CMC datasets and

curves and test the performance on a wider database.

Well known development of neural networks for

processing of massive image files can be easily used

in our method. The neural network is expected to

shorten the processing time significantly. We also

report the beginnings of a neural network based

recognition engine running on parallel GPU’s, which

is expected to enable real-time image recognition on

large databases. Finally, the recommended image

registration procedures are outlined which are

designed to optimize performance of the image

recognition algorithm by decreasing the number of

calculations necessary for image comparison.

REFERENCES

Maltoni, D., Maio D., Jain A. K., Prabhakar S., 2003,

Handbook of Fingerprint Recognition Springer-

Verlag, New York,

S. Pankanti, S. Prabhakar, A. K. Jain, CVPR 2001 Volume

1, 2001 pp I-805-I-812 vol.1. “On the individuality of

fingerprints” Proc of the 2001 IEEE Computer

Society Conference on Computer Vision and Pattern

Recognition.

Anil Jain, Arun Ross, Salil Prabhakar, 2001, pp. 282-285,

“Fingerprint Matching using Minutiae and Texture

Features” Proc. of international conference on Image

Processing.

Palma J, Liessner C, Mil’shtein S, 2006, “Contactless

Optical Scanning of Fingerprints with 180° View”

Scanning, 28, 6, pp 301-304

S. Mil’shtein, J. Palma, C. Liessner, M. Baier, A. Pillai,

and A.Shendye, 2008, "Line Scanner for Biometric

Applications" IEEE Intern. Conf. on Technologies for

Homeland Security, ISBN 978-1-4244-1978-4 P 205-

208.

P. Meenen, R. Adhami, 2005, “Approaches to image

binarization in current automated fingerprint

identification systems”, Proceedings of the Thirty-

Seventh Southeastern Symposium on System Theory,

ISBN: 0-7803-8808-9

C. Hu, J. Yin, E. Zhu, H. Chen, Y. Li, 2008, “Fingerprint

Alignment using Special Ridges” ISBN 978-1-4244-

2175-6

W. Chen, Y. Gao, 2007, “Minutiae-based Fingerprint

Alignment Using Phase Correlation”, Mechatronics

and Machine Vision in Practice, Springer Link, pp

193-198,

C. Carvalho, H. Yehia, 2004, “Fingerprint Alignment

using Line Segments”, Biometric Authentication,

Springer, pp 1-10

T. Zhang, J. Tian, Y. He, J. Cheng, X. Yang, 2003,

“Fingerprint alignment using similarity histogram”,

International conference on audio and video-based

biometric person authentication, pp 854-861

Xiaoxin Guo Zhiwen Xu Yinan Lu Yunjie Pang, Sep

2005, “An Application of Fourier-Mellin Transform in

Image Registration”, The Fifth International

Conference on Computer and Information

Technology, pp 619-623.

S. Prabhakar, A. Jain, “Fingerprint

Identification”,<http://www.cse.msu.edu/biometrics/fi

ngerprint.html>

Masters, Timothy, 1993, “Practical Neural Network

Recipies in C++” ISBN: 0-12-479040-2

B. Gour, T.K. Bandopadhyaya, R. Patel, 2010, “ART and

Modular Neural Network Architecture for multilevel

Categorization and Recognition of Fingerprints”,

2010 Third International Conference on Knowledge

Discovery and Data Mining, pp 536-539

A. L. H. Jin, A. Chekima, J. A. Dargham, L. C. Fan, 2002,

“Fingerprint Identification and Recognition using

Backpropagation Neural Network”, Proceeding of

Student Conference on Research and Development, pp

98-101

B. Gour, et. Al, 2005, “Fast Fingerprint Identification

System using Backpropagation Neural Network and

Self Organizing Map

”, Proc. Of Glow Gift.,

International Level Seminar.

VISAPP 2011 - International Conference on Computer Vision Theory and Applications

432