Pyramid-based Image Enhancement of Fingerprints

   

 

Halmstad University Post-Print

Pyramid-based Image Enhancement of Fingerprints

Hartwig Fronthaler, Klaus Kollreider and Josef Bigun

N.B.: When citing this work, cite the original article.

©2007 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

Fronthaler H, Kollreider K, Bigun J. Pyramid-based Image Enhancement of Fingerprints. In: 2007 IEEE Workshop on Automatic Identification Advanced Technologies proceedings : 7-8 June 2007, Alghero, Itlay: . IEEE; 2007. p. 45- 50.

DOI: http://dx.doi.org/10.1109/AUTOID.2007.380591 Copyright: IEEE

Post-Print available at: Halmstad University DiVA

http://urn.kb.se/resolve?urn=urn:nbn:se:hh:diva-2131

Pyramid-based Image Enhancement of Fingerprints

H. Fronthaler, K. Kollreider and J. Bigun Halmstad University, SE-301 18, Sweden

Abstract Whilenotacting against low quality fingerprints, methods to automatically assess the quality of a given impression, Reliablefeature extraction is crucial for accurate bio- such as [7, 12, 19], are useful and complementary to this metric recognition. Unfortunately feature extraction is study.

hampered by noisy input data, especially so in case offin- The task ofafingerprintenhancementalgorithmisto coun- gerprints. Wepropose a method to enhance the quality of teractthe aforesaid quality impairmentsandto reconstruct agivenfingerprint with the purpose to improve the recog- the actualfingerprintpattern as true tothe original aspos- nitionperformance. A Laplacian like image-scale pyramid sible. Furthermore, unrecoverable areas should be labeled is usedfor this purpose to decompose the original finger- as such, since fingerprint enhancement at too noisy parts print into 3 smaller images corresponding to

different

fre- yields spurious information. There are several published quency bands. In a further step, contextualfiltering is per- studiesonfingerprint imageenhancement. Hongetal. [13]

formed using these pyramid levels and ID Gaussians, where proposedan algorithm usingGaborband-passfilters tuned the corresponding filtering directions are derived from the tothecorresponding ridge frequency and orientationto re- frequency-adapted structuretensor. All image processing is moveundesired noise whilepreservingthetrueridge-valley done in the spatial domain, avoiding block artifacts while structures. Here, alloperationsareperformedinthespatial conserving the biometric signal well. We report on com- domain, whereas the contextual filtering in [8, 17] is done parative results and presentquantitative improvements, by inthe Fourier domain. Either way, block-wise processing applying the standardized NISTFIS2 fingerprint matcher to is usedto obtain the enhancement result causing restora- theFVC2004fingerprint database along with our as well as tion discontinuitiesatblock boundaries. These methodsare two other enhancements. The study confirms that the sug- likely successful in extreme bad quality regions, but also gested enhancement robustifiesfeature detection, e.g. minu- ratherrigidunder easy conditions. In [20,22], the finger- tiae, which in turn improves the recognition (20% relative print's block-wise power spectra weremultiplied bythem- improvementinequal error rate on DB3 ofFVC2004). selves but raised to the power of k, thus magnifying the dominant orientation. A block-wise Fourier transform is alsoemployed byChikkeruret al. [8, 9], followedby con- 1

Introduction

textualfiltering using raised cosines. Inarelatedstudy [2],a standarddiscrete scale space has been usedtocontextually process fingerprints. To the contrast, we employ an effi- Automatic fingerprint image enhancement plays a piv- cient multigridrepresentationof adiscretedifferential scale otal role in fingerprintrecognition, since it strongly deter- space and follow a different enhancement strategy. For a mines the success of all further steps. By succeeding the more detailedreview of fingerprint enhancement schemes sensing stageanddirectly preceding feature extraction, the we refer to [16].

purpose ofimageenhancement istofacilitate the latter by In this study wepropose the use of an image-scale pyramid

"denoising" the signal. It has been shownpreviously that and directionalfiltering in the spatial domain for fingerprint the quality ofa fingerprint image directly affects the per- image enhancement. Image pyramids or multi-resolution formance ofa given recognition system [11, 12, 15]. In processing is especially known from image compression anidealfingerprint image, ridges and valleys alternate and and medical image processing [10,14], but has not been uti- flow in alocally constant direction [16]. In realistic sce- lized to enhancefingerprint images before. The Laplacian nariosthough, the quality ofafingerprint imagemaysuffer pyramid [1, 18] resembles bandpass filteringin the spatial from variousimpairments, caused by i) scares and cuts, ii) domain. In this study, we decompose fingerprint images moist or dry skin, iii)sensor noise/blur, iv) wrong handling in a similar manner, since we expect all the relevant infor- of the sensor (e.g. too low/high contactpressure), v) gener- mation to be concentrated within a few frequency bands.

ally weak ridge-valley pattern of thegiven fingerprint, etc.

Furthermore, we propose Gaussian directional filtering to a)Pyramid DecompositionPD enhance theridge-valley pattern ofafingerprint imageus- Gaussian-like Laplacian-like ing computationally cheap ID filtering on higher pyramid

levels (lower resolution) only. Thefilteringdirections are 91 reduce(Th fo); 11=91-expand(92,f);

recovered from the orientations of the structure tensor [3] 92 =reduce

(gl,

f); 12

=92

- expand

(93,

f);

atthecorresponding pyramid level. Incontrast to other al-

gorithms, no block-wise processing is performed, thereby reduce (92, f); 13 93 - expand (94, f);

avoidingblockboundaryartifacts. 94 ⁼reduce(g3, f);

Inthe following section, adetailed description of the pro- b) Reconstruction R posed fingerprint enhancement algorithm is given. Insec- fp =expand (.,fo);

tion 3 wereportonexperimentsperformed on the FVC2004

expand

(., f) +

11

databaseusing the NIST FIS2 fingerprint matcher [21,23]. expand (13, f) + 12

2 Fingerprint Image Enhancement Table 1. The pyramid building process.

This section describes the proposed image enhancement_Thissecionescibesthepropsedimag ^enanceent of them from the nextlower

level, yielding 11-3.

The lat- forfingerprintsandpresents acomparativediscussion with

the fom

^the

nextelo leve, yediu

ⁿ

lo3 Thequat-

two other methods. The involved steps are arranged as

utraothed intfigur 1,ea voftem sto

^s

detailed

^{e below.} il- ter contain the adequately high, medium and low frequen-^cies

(ridge-valley pattern)

of the

original fingerprint.

It is worth

noting,

that

only

the

laplacian-like pyramid

levels 11-3 are used subsequently in this study. In a further step, [ -CE

OF

^L1 the contrast of the

single

^band

images

^isenhanced, follow-

originalt ' ^{Enhanced ing}

1i

^{=CE(li)where CE(x) = sign(x) x,to}

depre-

Fingerprint l--_____-._____ ---r;L-I-C Fingerprint

Fingerprint

ciate

^{small vectors ofx}

in comparison with

^{those oflarge}

P.D'

^{CE,OE R} Lmagnitudes.

1i

^ridge

--- pixels have negative values whereasvalley pixels areposi-

--CE,

^{OE- ---tive}

^(ideally).

^In

^figure

^2,an

example fingerprint

beside its

!___+.p lS I DF 13 --

_ x _contrast enhancedpyramidlevels 11-3 are displayed. The latterare used in aninitial reconstruction step R as shown rightmost. This reconstruction is crudesofar andrepresents

Fpige1d *imageso(ectangles)a

^onlyan

^isotropic (non-directional)

enhancement,

involving

api presg(o etnl s

signal

^values

starting

^at13 (see table 1). Itis alreadyvisi- ble that theportions of thefingerprint imagethat have been retained andcontrastimproved contain significant recogni- tioninformation, whereas otherscontaining high-frequency 2.1

Pyramid Decomposition (PD) isotropic noise are attenuated.

Worthmentioning, we cover approximately half the band- A pyramid decomposition requires resizing (scaling, or width of the

original image through

the

band-pass images

geometric transformation). To create our Gaussian and in total, e.g.

by setting fo

and

f

to 1.5. This choice de- Laplacian-like pyramids, we define the reduce

(I, f)

and pends onthe resolution of the

fingerprint

and

normally

^can expand

(I, f)

operations, which decrease and increase an be

approximated

off-line,either

experimentally

^or

by using

image I in size by the factorf,respectively. During reduce, informationonthe used ^sensor. In the used

images

studied the image is initially low-pass filtered to prevent aliasing here the

ridge frequency

ina

fingerprint image

is rzz.60cy- usingaGaussian kernel. The latter's standard deviation de- cles per

image width/height

[6]. This translatestoan

image

pends on theresizing factor, which here follows the lower dimension of

approximately

100-400

pixels

for the method bound approximation of the corresponding ideal low-pass to be most effective.

filter, ^{or =}

0?75

^f [3]. We initiallyreduce the original fin- gerprint image fp by a factor offo > 1.5 in orderto ex-

clude the

highest frequencies.

Ina further

step

we reduce 2.2 Orientation Estimation

(OE)

theimagesize byafactorf < 1.5 for three times. This is

also outlined to the left in table 1. To create images con-

taining only band limited signals of the original image, we The ridge-valley orientation for each of

11_3

is estimated expand the three images92-4 by factorf and subtract each using the complexstructuretensor approach [3]. The latter

Figure 2. i) Example fingerprint of the FVC2000-2 database, i-iv) its band-pass like decomposition via 11-3, v) the "so-far" reconstructed fingerprint.

tensoris builtasin

equation

1

z

[(D,G(uj) ^+jD G(jI)) llji

C.

[(x G(crj)

⁺

jy G(crj))

1*

l]2

(1)

where G((X1)

ecp

(-(x2+y2)/o1),j = ,and

"*" denotesa2D convolution. As theequation shows, the

operations

Dxli

and

Dyli

are realizedby means of convo- Figure 3. i-iii) HSV representation of LS1_3, lutions via

CxG(uil)

^{* 1i and}

CyG(uj1)

* 1i with C being steering V(magnitudes) and H (arguments).

the non-essential constant

-1/(72.

To obtain a robust es- timation of the dominant direction (linear symmetryorien- tation) at apoint, zi is averaged using a GaussianG

(u2),

where O2 > ^or to yield the complex image I20. Like- 2.3 Directional

Filtering

(DF) wise themagnitude ofzi is averaged to yield Ill. To be-

comeindependent of signalenergy, wecalculate

LSi

120 In orderto enhance the SNR (signal-to-noise ratio), i.e.

for level i, encoding local orientation (Z) and symmetry to remove sweatpores, scars, etc., weapply directional av- strength ( ). Also, by using LS, the magnitude of I20 is eragingtoall levels

l113.

Thefiltering direction within1iis attenuated if the underlying linear symmetry is not well- givenbyZ

(LSj)

/2-7/2,thus it follows theridges/valleys defined [4]. It is worth mentioning that all convolutions ofthefingerprint. Ateveryposition, the neighboring pixels areseparableandonlyID Gaussian(derivative) filters have along alinearemultiplied by a ID Gaussian and summed been used. Furthermore, LS1 is attenuated if its orienta- to yield the new value. The possible different line direc- tion deviates too much from theone ofLS2. This is done tionsarerestricted (here 20). Furthermore, wealsoexploit by LS1 ⁼ LS1 cos

(ZLS1

^-

ZLS2)

1. This is meaningful themagnitude of

LSi:

First, pixels where

ILS,l

< T, are because LS1 contains themostlocalized orientation (infor- assigned to the background, i.e. they are set to 0 (effec- mation also atminutia-level), but is also most susceptible tively amountingto asegmentation of the fingerprint from tonoise. Infigure3, LS1_3 for the example fingerprintare thebackgroundortheheavily noisy regions). Second, only displayed usingaHSVmodel,where itsmagnitudesmodu- if ILSi > T2 when measuredon asmall annulus centered late valueVand thearguments(local orientation) steerhue atthecurrentpixel, areasonablequality (presence of ridge- H. When comparedto a low-pass pyramid (e.g. Gaussian valleypattern atleveli)is ensured and the abovefiltering is pyramid), the estimated orientation inband-pass pyramids done. Otherwise, the pixel is again set to 0. In this way, fre- (e.g. Laplacian pyramid) was found much more robust, in quency selective structure tensors have helped tosmoothen

this context. the imageadaptively in the most appropriate direction per

Figure 4.

i-iii)

Directional filtered 11-3, iv) the reconstructed fingerprint, v) contrast enhancement.

layer. Atthe lowest level11, fine minutiaearepreserved be- afingerprint is done by puttingapre-determined threshold causethe LS1 filtering directions aresensitive tothem. At to ablock's energy. The final image in figure 5 depicts the higher levels 12_3 the rough ridge-valley flow is smoothed, result by the proposed method. As tobe expected, our ap- andgaps areclosed (e.g. caused by scars) because LS2-3 proach doesnotexhibit block-artifacts because the datapro- contain theglobal orientation. By use of the filtered levels cessing isnotblock-wise. Qualitatively, itappearsthatour 11-3 only, the image is reconstructed (R). A final contrast methodproduces highcontrastbetween ridges and valleys enhancement (CE) is done subsequently. In figure 4, the and the resultgenerally exhibits more fidelityto theorigi- filtered versions of11-3 for the example fingerprintaredis- nalcomparedtothe alternative approaches, because Hong's played, beside the reconstructed image and the final, con- methodappears somewhatmoreblurred while Chikkerur's trastenhancedimage. The ID Gaussians used for thispur- approach has visibleblock-artifacts, especiallynearminutia pose are small and thus singular points likecore and delta points. Being the basicresourceformostfingerprintrecog- points donotneed further attention. Theresulting enhanced nitiontechniques, including semiautomatic forensics, it isto fingerprints exhibitasmooth ridge-valley flow,yet preserv- beexpected that minituae neighbourhood degradation will ing the discriminative local and global information. decrease recognition performance. However,the validity of these qualitative observations needto be supportedexper- 2.4

Qualitative comparison and discussion imentally using publicly

available databases and standard

matching techniques,

which^we

present

^next.

Here, we presentsamples of enhancements toprovide a

visual feedback for a qualitative comparison between the 3

Experiments

proposed method and two other enhancement techniques

studied by Hong et al. [13] and Chikkerur et al. [8], re- In order to benchmark the capability of the proposed spectively. Implementations by the latter author, of both fingerprint enhancement algorithm, we need to test it methods were used. Figure 5 depicts a fingerprint image on highly corrupted fingerprint data, where reliable en- from the FVC2004-1 database together with 3 enhanced hancement becomes indispensable. Therefore we use the versionsasdeliveredby the mentioned techniques. Thesec- FVC2004 database [15], which was created to provide ondimage, correspondingtotheoutputof Hong'smethod, a tougher benchmark for state-of-the-art recognition sys- is achieved by theuse of Gabor-filters, tuned accordingto temsthanprevious fingerprint verification competitions [6].

estimatedfrequency and the orientation within small blocks Whencollectingthefingerprint data,individualswereasked of thefingerprint. The filtering is only performed if thecor- amongotherthings to vary thecontactpressure appliedto responding region exhibits ridge-valleystructurethat allows thesensorand theirfingerswereadditionallydriedormoist- correctenhancement. The result of Chikkerur's method is ened in orderto enforce challenging image quality condi- visualized in the thirdimage of figure 5. Here, all calcula- tions. Thetest setof the FVC2004 consists of4databases, tions areperformed in the frequency domain, using STFT whichwereacquired using different sensortypesand each (Short Time Fourier Transform), involving small overlap- of them contains 8 impressions of 100 fingers. Subse- ping blocks. The ridge frequency and orientation are de- quently, we will refer to these databases as DB1-4. It is termined in the Fourier domain, to steer the contextual fil- worth mentioning that DB4 was created using the SFinGE tering by steep band-pass functions. The segmentation of synthetic fingerprint generator [5] whereas DB 1-3 are pop-

Figure 5. i) Example fingerprint of FVC2004-1, ii) enhancement by Hong, iii) Chikkerur, iv) proposed.

ulatedby images representing authentic fingerprints sensed any of the useddatabases', neitherwere the others. Thus, by real sensors. When carryingoutfingerprint verification the availablefingerprint areaafter enhancement sometimes forasingleDB, wefollow the FVCprotocol involving 2800 happenedto be very small (especially in DB1). Infigure genuine trials and 4950 impostor trials. 6weshowexample fingerprints ofDB1-4 next totheiren- First of all,weenhance all fingerprints ofDB1-4with three hanced version, employing the proposed algorithm.

different enhancement methods: The proposed algorithm,

and the methods ofHongandChikkerur, whichwealready 4 Conclusion compared and inspected visually in section 2.4. Further-

more, weemployanindependent fingerprint matcher (NIST A novel image enhancement procedure for fingerprints FIS2 mindtct+ bozorth3

packages [21,23])

and take notes

. -

^.

of te ahieedER (qua Eror ate)perdatbas an has been presented. It

iS

a

continuous,

(in the sense of not of the achievedEER

(Equal

Error

Rate)

per databaseanda ^en lc-ie,saildmi prah tde hrb

enhancement method.

Ideally,

all of the latter should lead

ngtblock-wise), saidoin

^a

pproah Itsdoe treby

to lowererror rates for the fingerprint matcher,

compared

not suffer fromblocking artifacts. Both absolute frequency to when

matching

the

original fingerprints.

Table 2 shows

(isotropic information)

and orientation

(non-isotropic

in- formation) of the fingerprint pattern areutilized to obtain the enhancement. The former isimplemented by exploit- Enhancem. Method DB1 DB2

DB3

DB4 ing several levels of a band-pass

pyramid

and treating them

no

pre-enhancem. 14,5% 9,5%

M6,2% 7,3%

independently. The

typical

ridge-valley flowis coherence

nHong [13] (16,9%) 14,4% 7,1% 9,8% enhanced byusing directional averaging and the structure Chikkerur[8]

(19,1%)

11,9% 7,6% 10,9%

tensor direction

_causethemain

computations

at each level. The approacharedone at atleast 1.5is

efficient

times

be- proposed

method 12,0% 8,2%

15,0%

1 7,0% lower resolution than that of the

original,

^{andbyuse of}

ID

Table2. EER of the NIST FIS2 matcher on the filters only. The processing of the lowest level adds to the

original

and pre-enhanced FVC2004. fidelity and details (conservation of minutiae) whereas the rough ridge-valley flow is cleaned and gaps are closedat higher levels. Wehavecomparedourapproachto twoother enhancement methods, qualitatively and quantitatively by theEERof the matcher on all4FVC2004 databases: The use of a difficult fingerprint dataset. The results on the 4 results in the first row were achieved when using no im- FVC2004 databases are favorable to the suggested enhance- ageenhancementatall. The otherrowsdetail thematching ment method. While the alternative techniques have been performance if all impressions were initially enhanced by shown to improve the recognition performance in previous the method of Hong (second row), Chikkerur (third row) studies when processing fingerprints with more moderate and by the proposed approach (last row). Surprisingly, we noise than those affecting the FVC2004, the benefits of their can observe thatallenhancement methods buttheproposed enhancement apparently do not outweigh the introduced ar- worsen(!) theerrorrate. Our approach clearly leads tothe tifacts at presence ofheavy but realistic sensing noise.

lowest EER on all 4 databases. Worth mentioning, no pa-tOur algorithm was checked by inspection on selected fingerprints of rameters of our enhancement method have been adapted to DB2 from FVC2000(comparefigure 2-4)

N ~~References

[1] E.H.Adelson,C.H.Anderson,J. R.Bergen,P. J.Burt,and J. M.Ogden.

PyramiidMethods inImageProcessing.RCAEngineer, 29(6):33-41,1984.

[2] A.Almansa andT.Lindeberg. Fingerprintenhancementby shape adaptation ofscale-space operatorswith automatic scale-selection. IEEETransactions onImage Processing, 9(12):2027-2042,2000.

[3] J.Bigun. VisionwithDirection. Springer,2006.

[4] J.Bigun, ^H.Fronthaler, andK.Kollreider. Assuringliveness in biometric identity authenticationbyreal-time facetracking. InCIHSPS2004-IEEE International Conference^onComputationalIntelligence forHomeland Se- curityand PersonalSafety,Venice, Italy,pages 104-112.IEEECatalogNo.

04EX815,ISBN0-7803-8381-8,21-22July^2004.

[5] R.Cappelli,D.Maio, andD.Maltoni. Synthetic Fingerprint-ImageGenera- tion.InInternationalConference^onPatternRecognition,2000.

[6] R.Cappelli,D.Maio,D.Maltoni,J. L.Wayman,and A. K. Jain. Performance Evaluation ofFingerprintVerificationSystems.IEEE-PAMI,28(l):3-1 8,Jan-

uary206

[7] A. Chen, S.Dass, and A. Jain. Fingerprint Quality Indices forPredicting AuthenticationPerformance. InAudio- and Video-based Biometric Person Authentication(AVBPA) 2005,Rye Brook, NewYork,pages 160-170,July 2005.

[8] 5. Chkeu and V. Govindaraju. FingerprintImageEnhancementUsing STFTAnalysis.InInternationalWorkshoponPatternRecognitionfor^Crime Prevention,Securityand Surveillance(ICAPR05),pages20-29,2005.

[9] 5.Chikkerur,C.Wu,andV.Govindaraju.ASystematic Approachfor Feature Extraction inFingerprint Images. InInternationalConference^onBioinfor- maticsanditsApplications,pages344-350,2004.

[10] H. F. D.Kunz,K.Eck and T. Aach. A Nonlinear Multi-Resolution Gradient- AdaptiveFilter for MedicalImages.InSPIEMedicalImaging,volume5032, pages732-742,2003.

[11] J. Fierrez-Aguilar, L.-M. Munoz-Serrano, F. Alonso-Fernandez, and

J.Ortega-Garcia. Onthe effects ofimage quality degradationonminutiae- ii ~~~~~~~~~~and ridge-based automatic fingerprintrecognition. In IEEE Intl. Carnahan

[12] H.Fronthaler,K.Kollreider,and J.Bigun.AutomaticImageQualityAssess- Associationwith CVPR-06,NewYork,pages30-35,June2006.

[13] L.Hong,YWand,and A.Jain. Fingerprint imageenhancement: algorithm andperformanceevaluation.IEEE-PAMI,20(8):777-789,1998.

[14] D. Kaji. Improvement ofDiagnostic Image Quality Using a Frequency ProcessingBased on Decomposition into Multiresolution Space -Hybrid Processing-.Technicalreport,MISolutionGroup,2002.

[15] D.Maio,D.Maltoni,R.Cappelli,J.Wayman,and A.Jain. FVC2004: Third FingerprintVerificationCompetition.InInternationalConferenceonBiomet- ricAuthentication(ICBA04).HongKong,pages1-7,July2004.

[16] D.Maltoni,D.Maio,A. K.Jamn,and S. Prabhakar. Handbookoffingerprint recognition.Springer,2003.Includes DVD-ROM.

[17] B.G.Sherlock,D. M.Monro, andK.Millard. FingerprintEnhancementby directional FourierFiltering.InVisualImageSignal Processing,volume141, pages87-94,1994.

[18] E. P. Simoncelli and W. T. Freeman. The SteerablePyramid: AFlexible Architecture for Multi-Scale DerivativeComputation. InInternational Con- ferenceonImage Processing, volume3,pages444-447,23-26 Oct. 1995,

Washington,DC, USA,1995.

[19] E.Tabassi,C.Wilson,and C.Watson. Fingerprint ImageQuality. Technical Report NISTIR7151, Nist,2004.

[20] C.I. Watson,G.T.Candela,andP. J.Grother.ComparisonofFFTFingerprint FilteringMethods for Neural Network Classification. NISTIR,5493,1994.

[21] C.I. Watson,M. D.Garris, E. Tabassi,C.L.Wilson, R. M.McCabe,and S.Janet. Users Guid toFingerprintImage Softare2-NFIS2. NIST,2004.

UsewithLow-qualityPrints andDamaged Fingerprint.34:255-270,2001.

[23] HomePageof NIST:http://www.itl.nist.gov/div894/894.01/online.htm.

~~~~~~~~~~~~~~~~~~fc<~~~~~~~~~. .... ...

-~~~~~~~~~~~~~~~~~~~~~A~ ~...

Figure6.One~~~~~~~~~~~~~~~~~~~~~.finerrit.f.ac.D.o

FVC2004 next

to~~~~~~~~~~~

its

enhncd.erio

1-4244-1300-1/07/$25.OO 2007IEEE 50~~~~~~~~~~~~~~~~~~~~~~~~~~~~~...