xử lý vân tay

8/18/2019 xử lý vân tay

http://slidepdf.com/reader/full/xu-ly-van-tay 1/47

Fingerprint RecognitionBy

WUZHILI (99050056)

A thesis subitte! in p"rti"# $u#$i##ent o$

the re%uireents $or the !egree o$

B"che#or o$ &cience (Honors) in 'oputer &cience

'oputer &yste

Hong ong B"ptist Uniersity

*"te+ ,9 Apri# -00-

1



*ec#"r"tion

I Hereby declare that all the design and implementation of the

Honors Project is my independent effort except otherwise specified. I also

certify that this project have never been submitted for academic or

employment credit.

(99!!"# $u %hili

&omputer 'ystem ajor )epartment of &omputer 'cienceHong *ong +aptist ,niversity

)ate- 19/pril00

0



Hong ong B"ptist Uniersity

*ep"rtent o$ 'oputer &cience

'oputer &yste ."/or

$e hereby recommend that the Honors Project submitted by

$u%hili entitled 2ingerprint 3ecognition4 be accepted in partial

fulfillment of the re5uirements for the degree of +achelor of 'cience

(honors# in &omputer 'cience (&omputer 'ystem#.

'upervisor- )r. 6.6. 7ang 8bserver- )r 6.$. eung

)ate- )ate-

:



;



/c<nowledgement

I would li<e to express my great gratitude towards my supervisor= )r. 7ang=

6uan 6an who has given me much suggestion= support and help. /lso I would li<e to

than< my observer= )r. eung= 6iu wing= for his useful comments. I also ta<e this

opportunity to give than<s to all others who give me support for the project or in other

aspects of my study at H*+,.

!



Fingerprint Recognition+y

$,>HII (99!!"#

/bstract

y Honors Project is to investigate the current techni5ues for fingerprint

recognition. 7his target can be mainly decomposed into image preprocessing= feature

extraction and feature match. 2or each sub?tas<= some classical and up?to?date

methods in literatures are analy%ed. +ased on the analysis= an integrated solution for

fingerprint recognition is developed for demonstration.

y demonstration program is coded by /7/+. 2or the program= some

optimi%ation at coding level and algorithm level are proposed to improve the

performance of my fingerprint recognition system. 7hese performance enhancements

are shown by experiments conducted upon a variety of fingerprint images. /lso= the

experiments illustrate the <ey issues of fingerprint recognition that are consistent with

what the available literatures say.

"



7able of &ontents

&hapter 1- Introduction P11.1 $hat is / 2ingerprint P1

1.0 $hat is 2ingerprint 3ecognition P:1.: 7wo main approaches for 2ingerprint 3ecognition P!

&hapter 0- 'ystem )esign P"0.1 'ystem evel )esign P"0.0 /lgorithm evel )esign P@

&hapter :- 2ingerprint Image Preprocessing P9:.1 2ingerprint Image Anhancement P9:.0 2ingerprint Image +inari%ation P1:

:.: 2ingerprint Image 'egmentation P1;

&hapter ;- inutia Axtraction P1@;.1 2ingerprint 3idge 7hinning P1@;.0 inutia ar<ing P1B

&hapter ! inutia Post?processing P0!.1 2alse inutia 3emove P0!.0 ,nify inutia 3epresentation 2eature Cectors P00

&hapter "-inutia atch P0;".1 /lignment 'tage P0!".0 atch 'tage P0B

&hapter @- Axperimentation 3esults [email protected] Avaluation Indexes [email protected] Axperiment /nalysis P:

&hapter B- &onclusion P:0

3eferences P::

,ser anual P:!

@



&H/P77A3 1- ID738),&7I8D

1.1 $hat is / 2ingerprintE

/ fingerprint is the feature pattern of one finger (2igure 1.1.1#. It is believed with

strong evidences that each fingerprint is uni5ue. Aach person has his own fingerprints

with the permanent uni5ueness. 'o fingerprints have being used for identification and

forensic investigation for a long time.

2igure1.1.1 / fingerprint image ac5uired by an 8ptical 'ensor

/ fingerprint is composed of many ridges and furrows. 7hese ridges and furrows

present good similarities in each small local window= li<e parallelism and average

width.

However= shown by intensive research on fingerprint recognition= fingerprints are

not distinguished by their ridges and furrows= but by inutia= which are some

abnormal points on the ridges (2igure 1.1.0#. /mong the variety of minutia types

reported in literatures= two are mostly significant and in heavy usage- one is called

termination= which is the immediate ending of a ridgeF the other is called bifurcation=

which is the point on the ridge from which two branches derive.

B



2igure 1.1.0 inutia. (Calley is also referred as 2urrow=7ermination is also called Anding=

and +ifurcation is also called +ranch#

9



1.0 $hat is 2ingerprint 3ecognitionE

7he fingerprint recognition problem can be grouped into two sub?domains- one is

fingerprint verification and the other is fingerprint identification (2igure 1.0.1#. In

addition= different from the manual approach for fingerprint recognition by experts=

the fingerprint recognition here is referred as /23' (/utomatic 2ingerprint

3ecognition 'ystem#= which is program?based.

2igure 1.0.1 Cerification vs. Identification

2ingerprint verification is to verify the authenticity of one person by his fingerprint.

7he user provides his fingerprint together with his identity information li<e his I)

number. 7he fingerprint verification system retrieves the fingerprint template

according to the I) number and matches the template with the real?time ac5uired

fingerprint from the user. ,sually it is the underlying design principle of /2/'

(/utomatic 2ingerprint /uthentication 'ystem#.

2ingerprint identification is to specify one personGs identity by his fingerprint(s#.

$ithout <nowledge of the personGs identity= the fingerprint identification system tries

1



to match his fingerprint(s# with those in the whole fingerprint database. It is especially

useful for criminal investigation cases. /nd it is the design principle of /2I'

(/utomatic 2ingerprint Identification 'ystem#.

However= all fingerprint recognition problems= either verification or identification=

are ultimately based on a well?defined representation of a fingerprint. /s long as the

representation of fingerprints remains the uni5ueness and <eeps simple= the

fingerprint matching= either for the 1?to?1 verification case or 1?to?m identification

case= is straightforward and easy.

11



1.: 7wo approaches for 2ingerprint recognition

7wo representation forms for fingerprints separate the two approaches for

fingerprint recognition.

7he first approach= which is minutia?based= represents the fingerprint by its local

features= li<e terminations and bifurcations. 7his approach has been intensively

studied= also is the bac<bone of the current available fingerprint recognition products.

I also concentrate on this approach in my honors project.

7he second approach= which uses image?based methods"@= tries to do matching

based on the global features of a whole fingerprint image. It is an advanced and newly

emerging method for fingerprint recognition. /nd it is useful to solve some intractable

problems of the first approach. +ut my project does not aim at this method= so further

study in this direction is not expanded in my thesis.

10



&H/P7A3 0 '6'7A )A'IJD

0.1 'ystem evel )esign

/ fingerprint recognition system constitutes of fingerprint ac5uiring device= minutia

extractor and minutia matcher 2igure 0.1.1.

2igure 0.1.1 'implified 2ingerprint 3ecognition 'ystem

2or fingerprint ac5uisition= optical or semi?conduct sensors are widely used. 7hey

have high efficiency and acceptable accuracy except for some cases that the userGs

finger is too dirty or dry. However= the testing database for my project is from the

available fingerprints provided by 2C&00 (2ingerprint Cerification &ompetition

00#. 'o no ac5uisition stage is implemented.

7he minutia extractor and minutia matcher modules are explained in detail in the

next part for algorithm design and other subse5uent sections.

1:



0.0 /lgorithm evel )esign

7o implement a minutia extractor= a three?stage approach is widely used by

researchers. 7hey are preprocessing= minutia extraction and postprocessing stage

2igure 0.0.1.

2igure 0.0.1 inutia Axtractor

2or the fingerprint image preprocessing stage= I use Histogram A5uali%ation and

2ourier 7ransform to do image enhancement 9. /nd then the fingerprint image is

binari%ed using the locally adaptive threshold method 10. 7he image segmentation

tas< is fulfilled by a three?step approach- bloc< direction estimation= segmentation by

direction intensity ; and 3egion of Interest extraction by orphological operations.

ost methods used in the preprocessing stage are developed by other researchers but

they form a brand new combination in my project through trial and error. /lso the

morphological operations for extraction 38I are introduced to fingerprint image

segmentation by myself.

2or minutia extraction stage= three thinning algorithms 100 are tested and the

orphological thinning operation is finally bid out with high efficiency and pretty

1;



good thinning 5uality. 7he minutia mar<ing is a simple tas< as most literatures

reported but one special case is found during my implementation and an additional

chec< mechanism is enforced to avoid such <ind of oversight.

2or the postprocessing stage= a more rigorous algorithm is developed to remove

false minutia based on 101. /lso a novel representation for bifurcations is

proposed to unify terminations and bifurcations.

2igure0.0.0 inutia atcher

7he minutia matcher chooses any two minutia as a reference minutia pair and then

match their associated ridges first. If the ridges match well 1= two fingerprint images

are aligned and matching is conducted for all remaining minutia 2igure 0.0.0.

1!



&H/P7A3 : 2IDJA3P3ID7 I/JA P3AP38&A''IDJ

:.1 2ingerprint Image Anhancement

2ingerprint Image enhancement is to ma<e the image clearer for easy further

operations. 'ince the fingerprint images ac5uired from sensors or other medias are not

assured with perfect 5uality= those enhancement methods= for increasing the contrast

between ridges and furrows and for connecting the false bro<en points of ridges due

to insufficient amount of in<= are very useful for <eep a higher accuracy to fingerprint

recognition.

7wo ethods are adopted in my fingerprint recognition system- the first one is

Histogram A5uali%ationF the next one is 2ourier 7ransform.

1,1, Histogr" 2%u"#i3"tion+

Histogram e5uali%ation is to expand (mK rLng# the pixel value distribution (phMn

tNn# of an image so as to increase the perceptional information. 7he original histogram

of a fingerprint image has the bimodal type 2igure :.1.1.1= the histogram after the

histogram e5uali%ation occupies (nOm trong <hong# all the range from to 0!! and

the visuali%ation effect is enhanced 2igure :.1.1.0.

1"



2igure :.1.1 the 8riginal histogram

of a fingerprint image

2igure :.1.0 Histogram after the Histogram

A5uali%ation

7he right side of the following figure 2igure :.1.1.: is the output after the

histogram e5uali%ation.

2igure :.1.: Histogram Anhancement.

8riginal Image (eft#. Anhanced image (3ight#

1@



1,1- Fingerprint 2nh"nceent by Fourier 4r"ns$or

$e divide the image into small processing bloc<s (:0 by :0 pixels# and perform

the 2ourier transform according to-

(1#

for u Q = 1= 0= ...= :1 and v Q = 1= 0= ...= :1.

In order to enhance a specific bloc< by its dominant fre5uencies= we multiply the

227 of the bloc< by its magnitude a set of times. $here the magnitude of the original

227 Q abs(2(u=v## Q R2(u=v#R.

Jet the enhanced bloc< according to

(0# =

where 2?1(2(u=v## is done by-

(:#

for x Q = 1= 0= ...= :1 and y Q = 1= 0= ...= :1.

7he < in formula (0# is an experimentally determined constant= which we choose

<Q.;! to calculate. $hile having a higher S<S improves the appearance of the ridges=

filling up small holes in ridges= having too high a S<S can result in false joining of

ridges. 7hus a termination might become a bifurcation (TiUm <Vt thWc cX thU rY

nhNnh#. 2igure :.1.0.1 presents the image after 227 enhancement.

1B



2igure :.1.0.1 2ingerprint enhancement by 227Anhanced image (left#= 8riginal image (right#

7he enhanced image after 227 has the improvements to connect some falsely

bro<en points on ridges and to remove some spurious connections between ridges.

7he shown image at the left side of figure :.1.0.1 is also processed with histogram

e5uali%ation after the 227 transform. 7he side effect of each bloc< is obvious but it

has no harm to the further operations because I find the image after consecutive

binari%ation operation is pretty good as long as the side effect is not too severe.

19



:.0 2ingerprint Image +inari%ation

2ingerprint Image +inari%ation is to transform the B?bit Jray fingerprint image to a

1?bit image with ?value for ridges and 1?value for furrows. /fter the operation=

ridges in the fingerprint are highlighted with blac< color while furrows are white.

/ locally adaptive binari%ation method is performed to binari%e the fingerprint

image. 'uch a named method comes from the mechanism of transforming a pixel

value to 1 if the value is larger than the mean intensity value of the current bloc<

(1"x1"# to which the pixel belongs 2igure :.0.1.

2igure :.0.1 the 2ingerprint image after adaptive binari%ation

+inari%ed image(left#= Anhanced gray image(right#

0



:.: 2ingerprint Image 'egmentation

In general= only a 3egion of Interest (38I# is useful to be recogni%ed for each

fingerprint image. 7he image area without effective ridges and furrows is first

discarded since it only holds bac<ground information. 7hen the bound (biZn# of the

remaining effective area is s<etched (phNt h[a# out since the minutia in the bound

region are confusing with those spurious (gi# minutia that are generated when the

ridges are out of the sensor.

7o extract the 38I= a two?step method is used. 7he first step is bloc< direction

estimation (\]c l\^ng <h_i tr\`ng Tnh h\]ng# and direction variety chec< (<iUm tra

tr\`ng Tnh h\]ng# 1= while the second is intrigued from some orphological

methods.

1. B#oc !irection esti"tion

1.1 Astimate the bloc< direction for each bloc< of the fingerprint image with $x$

in si%e($ is 1" pixels by default#. 7he algorithm is-

I. &alculate the gradient values along x?direction (gx# and y?direction (gy# for

each pixel of the bloc<. 7wo 'obel filters are used to fulfill the tas<.

II. 2or each bloc<= use 2ollowing formula to get the east '5uare approximation

(xp x# of the bloc< direction.

tg0 Q 0 ∑ ∑ (gxgy#∑ ∑ (gx0?gy0# for all the pixels in each bloc<.

01



7he formula is easy to understand by regarding gradient values along x?

direction and y?direction as cosine value and sine value. 'o the tangent value

of the bloc< direction is estimated nearly the same as the way illustrated by the

following formula.

tg2θ = 2sinθ cosθ /(cos2θ -sin2θ )

1.0 /fter finished with the estimation of each bloc< direction= those bloc<s without

significant information on ridges and furrows are discarded (loi b# based on the

following formulas-

A Q 2 ∑ ∑ (gx*gy)+ ∑ ∑ (gx2-gy

2)}/ W*W*∑ ∑ (gx2+gy

2)

2or each bloc<= if its certainty level A is below a threshold (ng\ng#= then the

bloc< is regarded as a bac<ground bloc<.

7he direction map is shown in the following diagram. $e assume there is only onefingerprint in each image.

2igure :.:.1.1 )irection map.

+inari%ed fingerprint (left#= )irection map (right#

-1 RI e7tr"ction by .orpho#ogic"# oper"tions

00



7wo orphological operations called 8PADG and &8'AG are adopted. 7he

8PADG operation can expand images and remove pea<s introduced by bac<ground

noise 2igure :.:.0.0. 7he &8'AG operation can shrin< images and eliminate small

cavities 2igure :.:.0.:.

2igure :.:.0.1 8riginal Image /rea 2igure :.:.0.0 /fter &8'A operation

2igure :.:.0.: /fter 8PAD operation 2igure :.:.0.; 38I k +ound

2igure :.:.0.; show the interest fingerprint image area and its bound. 7he

bound is the subtraction of the closed area from the opened area. 7hen the algorithm

0:



throws away those leftmost= rightmost= uppermost and bottommost bloc<s out of the

bound so as to get the tightly bounded region just containing the bound and inner area.

0;



&H/P7A3 ; ID,7I/ A73/&7I8D

;.1 2ingerprint 3idge 7hinning

3idge 7hinning is to eliminate the redundant pixels of ridges till the ridges are just

one pixel wide. 10 uses an iterative= parallel thinning algorithm. In each scan of the

full fingerprint image= the algorithm mar<s down redundant pixels in each small

image window (:x:#. /nd finally removes all those mar<ed pixels after several scans.

In my testing= such an iterative= parallel thinning algorithm has bad efficiency

although it can get an ideal thinned ridge map after enough scans. 0 uses a one?in?

all method to extract thinned ridges from gray?level fingerprint images directly. 7heir

method traces along the ridges having maximum gray intensity value. However=

binari%ation is implicitly enforced since only pixels with maximum gray intensity

value are remained. /lso in my testing= the advancement of each trace step still has

large computation complexity although it does not re5uire the movement of pixel by

pixel as in other thinning algorithms. 7hus the third method is bid out which uses the

built?in orphological thinning function in /7/+.

7he thinned ridge map is then filtered by other three orphological operations to

remove some H brea<s= isolated points and spi<es.

0!



11

1

1

1

1

1 11

1

;.0 inutia ar<ing

/fter the fingerprint ridge thinning= mar<ing minutia points is relatively easy. +ut it

is still not a trivial tas< as most literatures declared because at least one special case

evo<es my caution during the minutia mar<ing stage.

In general= for each :x: window= if the central pixel is 1 and has exactly : one?value

neighbors= then the central pixel is a ridge branch 2igure ;.0.1. If the central pixel is

1 and has only 1 one?value neighbor= then the central pixel is a ridge ending

2igure;.0.0.

2igure ;.0.1 +ifurcation 2igure ;.0.0 7ermination

2igure ;.0.: 7riple counting branch

2igure ;.0.: illustrates a special case that a genuine branch is triple counted.

'uppose both the uppermost pixel with value 1 and the rightmost pixel with value 1

have another neighbor outside the :x: window= so the two pixels will be mar<ed as

0"



branches too. +ut actually only one branch is located in the small region. 'o a chec<

routine re5uiring that none of the neighbors of a branch are branches is added.

/lso the average inter?ridge width ) is estimated at this stage. 7he average inter?

ridge width refers to the average distance between two neighboring ridges. 7he way to

approximate the ) value is simple. 'can a row of the thinned ridge image and sum up

all pixels in the row whose value is one. 7hen divide the row length with the above

summation to get an inter?ridge width. 2or more accuracy= such <ind of row scan is

performed upon several other rows and column scans are also conducted= finally all

the inter?ridge widths are averaged to get the ).

7ogether with the minutia mar<ing= all thinned ridges in the fingerprint image are

labeled with a uni5ue I) for further operation. 7he labeling operation is reali%ed by

using the orphological operation- +$/+A.

0@



&H/P7A3 ! ID,7I/ P8'7P38&A''IDJ

!.1 2alse inutia 3emoval

7he preprocessing stage does not totally heal the fingerprint image. 2or example=

false ridge brea<s due to insufficient amount of in< and ridge cross?connections due to

over in<ing are not totally eliminated. /ctually all the earlier stages themselves

occasionally introduce some artifacts which later lead to spurious minutia. 7hese false

minutia will significantly affect the accuracy of matching if they are simply regarded

as genuine minutia. 'o some mechanisms of removing false minutia are essential to

<eep the fingerprint verification system effective.

'even types of false minutia are specified in following diagrams-

m1 m0 m: m;

m! m" m@

2igure !.1.1. 2alse inutia 'tructures. m1 is a spi<e piercing into a valley. In the m0case a spi<e falsely connects two ridges. m: has two near bifurcations located in the same ridge.7he two ridge bro<en points in the m; case have nearly the same orientation and a short distance.m! is ali<e the m; case with the exception that one part of the bro<en ridge is so short thatanother termination is generated. m" extends the m; case but with the extra property that a thirdridge is found in the middle of the two parts of the bro<en ridge. m@ has only one short ridgefound in the threshold window.

; only handles the case m1= m;=m! and m". 9 and 0 have not false minutia

removal by simply assuming the image 5uality is fairly good. 10 has not a

0B



systematic healing method to remove those spurious minutia although it lists all types

of false minutia shown in 2igure !.1.1 except the m: case.

y procedures in removing false minutia are-

1. If the distance between one bifurcation and one termination is less than ) and

the two minutia are in the same ridge(m1 case# . 3emove both of them. $here )

is the average inter?ridge width representing the average distance between two

parallel neighboring ridges.

0. If the distance between two bifurcations is less than ) and they are in the same

ridge= remove the two bifurcations. (m0= m: cases#.

:. If two terminations are within a distance ) and their directions are coincident

with a small angle variation. /nd they suffice the condition that no any other

termination is located between the two terminations. 7hen the two terminations

are regarded as false minutia derived from a bro<en ridge and are removed.

(case m;=m!= m"#.

;. If two terminations are located in a short ridge with length less than )= remove

the two terminations (m@#.

y proposed procedures in removing false minutia have two advantages. 8ne is

that the ridge I) is used to distinguish minutia and the seven types of false minutia arestrictly defined comparing with those loosely defined by other methods. 7he second

advantage is that the order of removal procedures is well considered to reduce the

computation complexity. It surpasses the way adopted by 10 that does not utili%e the

relations among the false minutia types. 2or example= the procedure: solves the m;=

m! and m" cases in a single chec< routine. /nd after procedure := the number of false

minutia satisfying the m@ case is significantly reduced.

09



1

11 1

!.0 ,nify terminations and bifurcations

'ince various data ac5uisition conditions such as impression pressure can easily

change one type of minutia into the other= most researchers adopt the unification

representation for both termination and bifurcation. 'o each minutia is completely

characteri%ed by the following parameters at last- 1# x?coordinate= 0# y?coordinate=

and :# orientation.

7he orientation calculation for a bifurcation needs to be specially considered. /ll

three ridges deriving from the bifurcation point have their own direction= 9

represents the bifurcation orientation using a techni5ue proposed in 1: 2igure

1.1.0. 1 simply chooses the minimum angle among the three anticloc<wise

orientations starting from the x?axis. +oth methods cast the other two directions away=

so some information loses. Here I propose a novel representation to brea< a

bifurcation into three terminations. 7he three new terminations are the three neighbor

pixels of the bifurcation and each of the three ridges connected to the bifurcation

before is now associated with a termination respectively 2igure !.0.1.

2igure !.0.1 / bifurcation to three terminations 7hree neighbors become terminations (eft#

Aach termination has their own orientation (3ight#

:



/nd the orientation of each termination (tx=ty# is estimated by following method

boo< ch0-

7rac< a ridge segment whose starting point is the termination and length is ). 'um

up all x?coordinates of points in the ridge segment. )ivide above summation with )

to get sx. 7hen get sy using the same way.

Jet the direction from- atan((sy?ty#(sx?tx##.

:1



&H/P7A3 " ID,7I/ /7&H

Jiven two set of minutia of two fingerprint images= the minutia match algorithm

determines whether the two minutia sets are from the same finger or not.

/n alignment?based match algorithm partially derived from the 1 is used in my

project. It includes two consecutive stages- one is alignment stage and the second is

match stage.

1. /lignment stage. Jiven two fingerprint images to be matched= choose any oneminutia from each image= calculate the similarity of the two ridges associated

with the two referenced minutia points. If the similarity is larger than a

threshold= transform each set of minutia to a new coordination system whose

origin is at the referenced point and whose x?axis is coincident with the direction

of the referenced point.

0. atch stage- /fter we get two set of transformed minutia points= we use the

elastic match algorithm to count the matched minutia pairs by assuming two

minutia having nearly the same position and direction are identical.

:0



".1 /lignment 'tage

1. 7he ridge associated with each minutia is represented as a series of x?

coordinates (x1= x0xn# of the points on the ridge. / point is sampled per ridge length

starting from the minutia point= where the is the average inter?ridge length. /nd n

is set to 1 unless the total ridge length is less than 1.

'o the similarity of correlating the two ridges is derived from-

' Q ∑miQxii∑m

iQxi0i

0.!=

where (xixn# and (i D # are the set of minutia for each fingerprint image

respectively. /nd m is minimal one of the n and D value. If the similarity score is

larger than .B= then go to step 0= otherwise continue to match the next pair of ridges.

0. 2or each fingerprint= translate and rotate all other minutia with respect to the

reference minutia according to the following formula-

xinew

yinew

θinew

xi x−( #

yi y−( #

θi

θ−( )

=TM *

=where (x=y=θ# is the parameters of the reference minutia= and 7 is

TM =

cosθ

sinθ

sinθ−

cosθ

1

7he following diagram illustrate the effect of translation and rotation-

::



θ

)

A

2

?axis

6?axis

q?axis

6q?axis

y

xθ

)

A

2

θ

)

A

2

)

A

2

?axis

6?axis

q?axis

6q?axis

y

x

7he new coordinate system is originated at minutia 2 and the new x?axis is coincident with thedirection of minutia 2. Do scaling effect is ta<en into account by assuming two fingerprints from thesame finger have nearly the same si%e.

y method to align two fingerprints is almost the same with the one used by 1 but

is different at step 0. inGs method uses the rotation angle calculated from all the

sparsely sampled ridge points. y method use the rotation angle calculated earlier by

densely tracing a short ridge start from the minutia with length ). 'ince I have

already got the minutia direction at the minutia extraction stage= obviously my method

reduces the redundant calculation but still holds the accuracy.

/lso inGs way to do transformation is to directly align one fingerprint image to

another according to the discrepancy of the reference minutia pair. +ut it still re5uires

a transform to the polar coordinate system for each image at the next minutia match

:;



stage. y approach is to transform each according to its own reference minutia and

then do match in a unified x?y coordinate. 7herefore= less computation wor<load is

achieved through my method.

:!



".0 atch 'tage

7he matching algorithm for the aligned minutia patterns needs to be elastic since the

strict match re5uiring that all parameters (x= y= θ# are the same for two identical

minutia is impossible due to the slight deformations and inexact 5uanti%ations of

minutia.

y approach to elastically match minutia is achieved by placing a bounding box

around each template minutia. If the minutia to be matched is within the rectangle box

and the direction discrepancy between them is very small= then the two minutia are

regarded as a matched minutia pair. Aach minutia in the template image either has no

matched minutia or has only one corresponding minutia.

7he final match ratio for two fingerprints is the number of total matched pair

over the number of minutia of the template fingerprint. 7he score is 1ratio and

ranges from to 1. If the score is larger than a pre?specified threshold= the two

fingerprints are from the same finger.

However= the elastic match algorithm has large computation complexity and is

vulnerable to spurious minutia.

:"



&H/P7A3 @ APA3IAD7/7I8D 3A',7'

@.1 Avaluation indexes for fingerprint recognition

7wo indexes are well accepted to determine the performance of a fingerprint

recognition system- one is 233 (false rejection rate# and the other is 2/3 (false

acceptance rate#. 2or an image database= each sample is matched against the

remaining samples of the same finger to compute the 2alse 3ejection 3ate. If the

matching g against h is performed= the symmetric one (i.e.= h against g # is not

executed to avoid correlation. /ll the scores for such matches are composed into a

series of &orrect 'core. /lso the first sample of each finger in the database is matched

against the first sample of the remaining fingers to compute the 2alse /cceptance

3ate. If the matching g against h is performed= the symmetric one (i.e.= h against g # is

not executed to avoid correlation. /ll the scores from such matches are composed into

a series of Incorrect 'core.

:@



@.0 Axperimentation 3esults

/ fingerprint database from the 2C&0 (2ingerprint Cerification &ompetition

0# is used to test the experiment performance. y program tests all the images

without any fine?tuning for the database. 7he experiments show my program can

differentiate imposturous minutia pairs from genuine minutia pairs in a certain

confidence level. 2urthermore= good experiment designs can surely improve the

accuracy as declared by 1. 2urther studies on good designs of training and testing

are expected to improve the result.

Here is the diagram for &orrect 'core and Incorrect 'core distribution-

2igure @.0.1 )istribution of &orrect 'cores and Incorrect 'cores3ed line- Incorrect 'coreJreen line- &orrect 'cores

It can be seen from the above figure that there exist two partially overlapped

distributions. 7he 3ed curve whose pea<s are mainly located at the left part means the

average incorrect match score is 0!. 7he green curve whose pea<s are mainly located

on the right side of red curve means the average correct match score is :!. 7his

:B



indicates the algorithm is capable of differentiate fingerprints at a good correct rate by

setting an appropriate threshold value.

2igure @.0.0 2/3 and 233 curve+lue dot line- 233 curve3ed dot line- 2/3 curve

7he above diagram shows the 233 and 2/3 curves. /t the e5ual error rate 0!= the

separating score :: will falsely reject 0! genuine minutia pairs and falsely accept

0! imposturous minutia pairs and has @! verification rate.

7he high incorrect acceptance and false rejection are due to some fingerprint images

with bad 5uality and the vulnerable minutia match algorithm.

:9



&H/P7A3 B &8D&,'7I8D

y project has combined many methods to build a minutia extractor and a minutiamatcher. 7he combination of multiple methods comes from a wide investigation into

research paper. /lso some novel changes li<e segmentation using orphological

operations= minutia mar<ing with special considering the triple branch counting=

minutia unification by decomposing a branch into three terminations= and matching in

the unified x?y coordinate system after a two?step transformation are used in my

project= which are not reported in other literatures I referred to.

/lso a program coding with /7/+ going through all the stages of the

fingerprint recognition is built. It is helpful to understand the procedures of

fingerprint recognition. /nd demonstrate the <ey issues of fingerprint recognition.

;



3A2A3AD&A'

1 in Hong. S/utomatic Personal Identification ,sing 2ingerprintsS= Ph.). 7hesis=

199B.0 ).aio and ). altoni. )irect gray?scale minutiae detection in fingerprints.

IAAA 7rans. Pattern /nal. /nd achine Intell.= 19(1#-0@?;= 199@.

: ain= /.*.= Hong= .= and +olle= 3.(199@#= 8n?ine 2ingerprint Cerification=4

IAAA 7rans. 8n Pattern /nal and achine Intell= 19(;#= pp. :0?:1;.

; D. 3atha= '. &hen and /.*. ain= S/daptive 2low 8rientation +ased 2eature

Axtraction in 2ingerprint ImagesS= Pattern 3ecognition= Col. 0B= pp. 1"!@?1"@0=

Dovember 199!.

! /lessandro 2arina= >solt .*ovacs?Cajna= /lberto leone= 2ingerprint minutiae

extraction from s<eletoni%ed binary images= Pattern 3ecognition= Col.:0= Do.;=

ppB@@?BB9= 1999.

" ee= &..= and $ang= '.).- 2ingerprint feature extration using Jabor filters=

Alectron. ett.= 1999= :!= (;#= pp.0BB?09.

@ . 7ico= P.*uosmanen and .'aarinen. $avelet domain features for fingerprint

recognition= Alectroni. ett.= 01= :@= (1#= pp.01?00.

B . Hong= 6. $an and /.*. ain= S2ingerprint Image Anhancement- /lgorithms

and Performance AvaluationS= IAAA 7ransactions on P/I =Col. 0= Do. B= pp.@@@?

@B9= /ugust 199B.

9 Image 'ystems Angineering Program= 'tanford ,niversity. 'tudent project +y

7homas 6eo= $ee Peng 7ay= 6ing 6u 7ai.

http-ise.stanford.educlassee:"Baproj1dropboxproject00finger

1 2C&0. http-bias.csr.unibo.itfvc0

11 2C&00. http-bias.csr.unibo.itfvc00

;1

http://ise0.stanford.edu/class/ee368a_proj01/dropbox/project22/finger/

http://bias.csr.unibo.it/fvc2000/




http://ise0.stanford.edu/class/ee368a_proj01/dropbox/project22/finger/



10 .&. ain= ,.Halici= I. Hayashi= '.+. ee and '.7sutsui. Intelligent biometric

techni5ues in fingerprint and face recognition. 1999= the &3& Press.

1: . . )onahue and '. I. 3o<hlin= S8n the ,se of evel &urves in Image

/nalysis=S Image ,nderstanding= C8. !@= pp "!0 ? "!!= 1990.

;0



User ."nu"#

1. 7ype command st"rt8gui8sing#e8o!e in /7/+ ".

2igure .1 the ,ser Interface of the 2ingerprint 3ecognition 'ystem. 7he series of buttons on theleft side will be invo<ed se5uentially in the conse5uent demonstration. 7he two blan< areas are

used to show the fingerprint image before and after a transaction respectively.

0. &lic< Lo"! +utton

2igure .0 oad a gray level fingerprint image from a drive specified by ,sers. ultiple formats

are supported and the image si%e is not limited. +ut the fingerprint ridges should have large grayintensity comparing with the bac<ground and valleys.

;:



:. &lic< his2%u"#i3"tion +utton

2igure .: /fter Histogram A5uali%ation. 7he image on the left side is the original fingerprint.7he enhanced image after the Histogram A5uali%ation is shown on the right side.

;. &lic< $$t +utton

2igure .; &aptured window after clic< 227G button. 7he pop?up dialog accepts the parameter < (please refer the formula 0#. 7he experimental optimal < value is .;!. ,sers can fill any other constant in the dialog to get a better performance. 7he enhanced image will be shown in the leftscreen box= which however is not shown here.

!. &lic< Bin"ri3"tion +utton

".&lic< *irection +utton

;;



2igure .!." 'creen capture after binari%ation (left# and bloc< direction estimation (right#.

@. &lic< RI Are" +utton

2igure .@ 38I extraction(right#. 7he intermediate steps for all the morphological operations suchclose and open are not shown. 7he right screen box shows the final region of interest of thefingerprint image. 7he subse5uent operations will only operate on the region of interest.

B. &lic< 4hinning +utton

9. &lic< Reoe H bre"s +utton

1. &lic< Reoe spies +utton

;!



2igure .1 the 2ingerprint image after thining= H brea<s removal= isolated pea<s removal andspi<e removal.(right#.

11. &lic< 27tr"ct +utton

10. &lic< Re"# .inuti" +utton

2igure .10 inutia ar<ing (right# and 2alse inutia 3emoval (eft#. +ifurcations are locatedwith yellow crosses and terminations are denotes with red stars. /nd the genuine minutia (left# arelabeled with orientations with green arrows.

1:. &lic< &"e +utton

;"



2igure .1: 'ave minutia to a text file. 7he saved text file stores the information on all genuineminutia. 7he exact format of the files are explained in the source code.

1;. &lic< ."tch +utton

2igure .1; oad two minutia files and do matching. ,sers can open two minutia data files fromthe dialog invo<ed by clic<ing the atchG button. 7he match algorithm will return a prompt of thematch score. +ut be noted that matching in the J,I mode is not encouraged since the matchalgorithm relies on heavy computation. ,npredicted states will happen after a long irresponsiverunning time. +atch testing is prepared for testing match. Please refer the source files for batchtesting.

Documents

xử lý vân tay