13
A modified Gabor filter design method for fingerprint image enhancement Jianwei Yang, Lifeng Liu, Tianzi Jiang * , Yong Fan * National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, P.O. Box 2728, Beijing 100080, PR China Received 17 June 2002; received in revised form 8 January 2003 Abstract Fingerprint image enhancement is an essential preprocessing step in fingerprint recognition applications. In this paper, we propose a novel filter design method for fingerprint image enhancement, primarily inspired from the tra- ditional Gabor filter (TGF). The previous fingerprint image enhancement methods based on TGF banks have some drawbacks in their image-dependent parameter selection strategy, which leads to artifacts in some cases. To address this issue, we develop an improved version of the TGF, called the modified Gabor filter (MGF). Its parameter selection scheme is image-independent. The remarkable advantages of our MGF over the TGF consist in preserving fingerprint image structure and achieving image enhancement consistency. Experimental results indicate that the proposed MGF enhancement algorithm can reduce the FRR of a fingerprint matcher by approximately 2% at a FAR of 0.01%. Ó 2003 Published by Elsevier Science B.V. Keywords: Fingerprints; Enhancement; Traditional Gabor filter; Modified Gabor filter; Parameter selection; Low pass filter; Band pass filter 1. Introduction Fingerprint recognition is being widely applied in the personal identification for the purpose of high degree of security. However, some fingerprint images captured in variant applications are poor in quality, which corrupts the accuracy of fingerprint recognition. Consequently, fingerprint image en- hancement is usually the first step in most auto- matic fingerprint identification systems (AFISs). There have existed a variety of research activi- ties along the stream of reducing noises and in- creasing the contrast between ridges and valleys in the gray-scale fingerprint images. Some ap- proaches are implemented in spatial domain, others in frequency domain. OÕGorman and Nickerson (1989) and Mehtre (1993) performed fingerprint image enhancement based on directional filters; Maio and Maltoni (1998) employed neural network in minutiae filtering; Almansa and Lindeberg (2000) enhanced them in scale space; * Corresponding authors. Tel.: +86-10-8261-4469; fax: +86- 10-6255-1993. E-mail addresses: [email protected] (J. Yang), lfliu@ nlpr.ia.ac.cn (L. Liu), [email protected] (T. Jiang), yfan@ nlpr.ia.ac.cn (Y. Fan). 0167-8655/03/$ - see front matter Ó 2003 Published by Elsevier Science B.V. doi:10.1016/S0167-8655(03)00005-9 Pattern Recognition Letters xxx (2003) xxx–xxx www.elsevier.com/locate/patrec ARTICLE IN PRESS

PRL_Yang

Embed Size (px)

DESCRIPTION

Fingerprint image enhancement is an essential preprocessing step in fingerprint recognition applications. In thispaper, we propose a novel filter design method for fingerprint image enhancement, primarily inspired from the traditionalGabor filter (TGF). The previous fingerprint image enhancement methods based on TGF banks have somedrawbacks in their image-dependent parameter selection strategy, which leads to artifacts in some cases. To address thisissue, we develop an improved version of the TGF, called the modified Gabor filter (MGF). Its parameter selectionscheme is image-independent. The remarkable advantages of our MGF over the TGF consist in preserving fingerprintimage structure and achieving image enhancement consistency. Experimental results indicate that the proposed MGFenhancement algorithm can reduce the FRR of a fingerprint matcher by approximately 2% at a FAR of 0.01%. 2003 Published by Elsevier Science B.V.

Citation preview

  • A modified Gabor filter design method for fingerprintimage enhancement

    Jianwei Yang, Lifeng Liu, Tianzi Jiang *, Yong Fan *

    National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, P.O. Box 2728,

    Beijing 100080, PR China

    Received 17 June 2002; received in revised form 8 January 2003

    Abstract

    Fingerprint image enhancement is an essential preprocessing step in fingerprint recognition applications. In this

    paper, we propose a novel filter design method for fingerprint image enhancement, primarily inspired from the tra-

    ditional Gabor filter (TGF). The previous fingerprint image enhancement methods based on TGF banks have some

    drawbacks in their image-dependent parameter selection strategy, which leads to artifacts in some cases. To address this

    issue, we develop an improved version of the TGF, called the modified Gabor filter (MGF). Its parameter selection

    scheme is image-independent. The remarkable advantages of our MGF over the TGF consist in preserving fingerprint

    image structure and achieving image enhancement consistency. Experimental results indicate that the proposed MGF

    enhancement algorithm can reduce the FRR of a fingerprint matcher by approximately 2% at a FAR of 0.01%.

    2003 Published by Elsevier Science B.V.

    Keywords: Fingerprints; Enhancement; Traditional Gabor filter; Modified Gabor filter; Parameter selection; Low pass filter; Band pass

    filter

    1. Introduction

    Fingerprint recognition is being widely appliedin the personal identification for the purpose of

    high degree of security. However, some fingerprint

    images captured in variant applications are poor in

    quality, which corrupts the accuracy of fingerprint

    recognition. Consequently, fingerprint image en-

    hancement is usually the first step in most auto-

    matic fingerprint identification systems (AFISs).There have existed a variety of research activi-

    ties along the stream of reducing noises and in-

    creasing the contrast between ridges and valleys

    in the gray-scale fingerprint images. Some ap-

    proaches are implemented in spatial domain, others

    in frequency domain. OGorman and Nickerson(1989) and Mehtre (1993) performed fingerprint

    image enhancement based on directional filters;Maio and Maltoni (1998) employed neural

    network in minutiae filtering; Almansa and

    Lindeberg (2000) enhanced them in scale space;

    *Corresponding authors. Tel.: +86-10-8261-4469; fax: +86-

    10-6255-1993.

    E-mail addresses: [email protected] (J. Yang), lfliu@

    nlpr.ia.ac.cn (L. Liu), [email protected] (T. Jiang), yfan@

    nlpr.ia.ac.cn (Y. Fan).

    0167-8655/03/$ - see front matter 2003 Published by Elsevier Science B.V.doi:10.1016/S0167-8655(03)00005-9

    Pattern Recognition Letters xxx (2003) xxxxxx

    www.elsevier.com/locate/patrec

    ARTICLE IN PRESS

    mail to: [email protected]

  • Greenberg et al. (2000) and Jiang (2001) resorted

    to an anisotropic filter and an oriented low pass

    filter to suppress noises respectively. In contrast

    with the above methods in spatial domain, Sher-

    lock et al. (1994), Willis and Myers (2000) and

    Kamei and Mizoguchi (1995) denoised fingerprintimages in frequency domain. There are advantages

    and disadvantages of analysis merely in spatial

    domain or frequency domain. As is well known,

    the Gabor filter is a very useful tool for texture

    analysis in both domains and hence combines the

    advantages of both filters. Considering their fre-

    quency-selective and orientation-selective proper-

    ties and optimal joint resolution in both domains,Hong et al. (1998) made use of Gabor filter banks

    to enhance fingerprint images and reported to

    achieve good performance.

    In their algorithm, called the traditional Gabor

    filter (TGF) method in this paper, Hong et al.

    assumed that the parallel ridges and valleys exhibit

    some ideal sinusoidal-shaped plane waves associ-

    ated with some noises. In other words, the 1-Dsignal orthogonal to the local orientation is ap-

    proximately a digital sinusoidal wave. Then, the

    TGF is tuned to the corresponding local orienta-

    tion and ridge frequency (reciprocal of ridge dis-

    tance) in order to remove noises and preserve the

    genuine ridge and valley structures. Unfortunately,

    their prior sinusoidal plane wave assumption is

    inaccurate because the signal orthogonal to the

    local orientation in practice does not consist of

    an ideal digital sinusoidal plane wave in some

    fingerprint images or some regions (see Fig. 1).

    Moreover, the TGFs parameter selection in theirmethod (such as the standard deviation of the

    Gaussian function) is empirical. This implemen-

    tation implies the disadvantage of image-depen-

    dence. In some cases, it could unexpectedly result

    in inconsistent image enhancement, which is

    baneful to the following steps.

    In order to overcome its shortcomings, we

    improve the TGF to the modified Gabor filter(MGF) by discarding the inaccurate prior sinu-

    soidal plane wave assumption. Our MGFs pa-rameters are deliberately specified through some

    principles instead of experience and an image-

    independent parameter selection scheme is ap-

    plied. Experimental results illustrate that our

    MGF could achieve better performance than the

    TGF.The rest of this paper is organized as follows. In

    Section 2, we briefly introduce the TGF and our

    MGF, and meanwhile highlight our motivation of

    extending the TGF to the MGF. Section 3 is de-

    voted to the parameter selection of our MGF.

    Implementation is detailed in Section 4. Experi-

    Fig. 1. A fingerprint image and corresponding ridge and valley topography. The top-right region can be approximately treated as a

    sinusoidal plane wave, but never the bottom-left.

    2 J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx

    ARTICLE IN PRESS

  • mental results are shown in Section 5. In Section 5

    we made some conclusions.

    2. Traditional Gabor filter and modified Gabor filter

    The Gabor function has been recognized as a

    very useful tool in computer vision and image

    processing, especially for texture analysis, due to

    its optimal localization properties in both spatial

    and frequency domain. There are lots of papers

    published on its applications since Gabor (1946)

    proposed the 1-D Gabor function. The family of

    2-D Gabor filters was originally presented byDaugman (1980) as a framework for understand-

    ing the orientation-selective and spatialfrequency-

    selective receptive field properties of neurons in the

    brains visual cortex, and then was further mathe-matically elaborated (Daugman, 1985).

    The 2-D Gabor function is a harmonic oscil-

    lator, composed of a sinusoidal plane wave of a

    particular frequency and orientation, within aGaussian envelope. A complex 2-D Gabor filter

    over the image domain x; y is defined as

    Gx; y exp

    x x02

    2r2x y y0

    2

    2r2y

    !

    exp2piu0x x0 v0y y0 1

    where x0; y0 specify the location in the image,u0; v0 specifying modulation that has spatialfrequency x0

    ffiffiffiffiffiffiffiffiffiffiffiffiffiffiu20 v20

    pand orientation h0

    arctanv0=u0, and rx and ry are the standard de-viations of the Gaussian envelope respectively

    along x-axis and y-axis. Derived from formula (1)by elaborately selecting above parameters, the

    even-symmetric real component of the original 2-D Gabor filer can be obtained, which is adopted in

    (Jain and Farrokhnia, 1991; Hong et al., 1998):

    gx; y; T ;/ exp

    12

    x2/r2x

    "

    y2/r2y

    #!cos

    2px/T

    2

    x/ x cos/ y sin/ 3

    y/ x sin/ y cos/ 4

    where / is the orientation of the derived Gaborfilter, and T is the period of the sinusoidal planewave.

    If we decompose formula (2) into two ortho-

    gonal parts, one parallel and the other perpendi-cular to the orientation /, the following formulacan be deduced:

    gx; y; T ;/ hxx; T ;/ hyy;/

    exp (

    x2/2r2x

    !cos

    2px/T

    )

    exp (

    y2/2r2y

    !)5

    The first part hx behaves as a 1-D Gabor functionwhich is a band pass filter, and the second one hyrepresents a Gaussian function which is a low pass

    filter. Therefore, a 2-D even-symmetric Gabor fil-

    ter (TGF) performs a low pass filtering along

    the orientation / and a band pass filtering or-thogonal to its orientation /. The band pass andlow pass properties along the two orthogonal

    orientations are very beneficial to enhancing fin-

    gerprint images, since these images usually show a

    periodic alternation between ridges and valleys

    orthogonal to the local orientation and parallel

    exhibit an approximate continuity along the local

    orientation.

    It should be pointed out that hx in formula (5)could be treated as a non-admissible mother

    wavelet (indicated by its Fourier representation

    hhx0 6 0). Its band pass property is related withthe rx. If rx is too small, the band pass filter de-generates into a low pass function (indicated by its

    Fourier representation hhx0 0). On the otherhand, if rx is appropriately large, hx can be ap-proximately regarded as an admissible motherwavelet (indicated by its Fourier representation

    hhx0 0) with good band pass property (see Fig.2).

    For the purpose of enhancing fingerprint images

    by the TGF, Hong et al. (1998) assumed that ridges

    and valleys show a sinusoidal plane wave pattern

    and specified the parameter T in formula (2) or (5)as the distance between two successive ridges.However, in practice this prior assumption is in-

    accurate. In Fig. 1, the ridge and valley structures

    J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx 3

    ARTICLE IN PRESS

  • in the top-right region can be roughly regarded as a

    sinusoidal plane wave pattern, but never those in

    the bottom-left region. In that case, the TGF

    method fails. This phenomenon can be explicitly

    explained in frequency domain. Although a band

    pass filter can enlarge the signal of a particular

    frequency and suppress others, the preferred fre-

    quency cannot be accurately specified in somecases. In other words, the ridge and valley pattern

    like the bottom-left region of Fig. 1 is not com-

    posed of a sinusoidal plane wave of only a partic-

    ular frequency but a periodic one whose Fourier

    extension contains different frequency harmonics.

    The TGF cannot pass the entire harmonics except

    the signal of a particular frequency. However, the

    low frequency components usually contain usefultexture information (e.g. slow variation of intensi-

    ties near ridges centers orthogonal to the localorientation is represented as low frequency com-

    ponents, see Fig. 1(b)). Thereby, the TGF method

    may lose some useful original information.

    To overcome the TGFs drawbacks mentionedabove, we replace the cosine function cosx; T informula (2) and (5) with another periodic function

    F x; T1; T2 to construct our MGF. It is incorpo-rated with two cosinusoidal functional curves with

    different periods T1 and T2 (see Fig. 3). The partsabove the x-axis consist of a cosinusoidal func-tional curve with a period T1 and the ones belowthe x-axis consist of another cosinusoidal func-tional curve with different period T2. F x; T1; T2 isextended periodically and elaborated mathemati-

    cally as follows:

    F x;T1;T2 f x

    xT1=2 T2=2

    T1=2 T2=2

    6

    Fig. 2. The TGF and its response represented in spatial and frequency domain.

    4 J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx

    ARTICLE IN PRESS

  • f x

    cos 2pxT1

    06x6T1=4

    cos 2pxT1=4T2=4T2

    T1=4< x< T1=4T2=2

    cos 2pxT1=2T2=2T1

    T1=4T2=26x6T1=2T2=2

    8>>>>>>>>>:

    7

    where btc floort means the largest integer notlarger than t.

    From the above definition, F x; T1; T2 is a pe-riodic even-symmetric oscillator with the period

    T1 T2=2, and it becomes a true cosinusoidalfunction when T1 T2. Then, our MGF can bespecified by modulating the periodic function

    F x; T1; T2 by a 2-D anisotropic Gaussian func-tion. So, formula (5) is turned into

    Fig. 3. The periodic function F x; T1; T2. The parts above thex-axis consist of a cosinusoidal functional curve with period T1,and those below the x-axis consist of another cosinusoidalfunctional curve with different period T2.

    Fig. 4. Our MGF and its response represented in spatial and frequency domain.

    J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx 5

    ARTICLE IN PRESS

  • g0x; y; T1; T2;/ h0xx; T1; T2;/ h0yy;/

    exp (

    x2/2r2x

    !F x/; T1; T2

    )

    exp (

    y2/2r2y

    !)8

    The frequency representation of the MGF is nolonger a band pass filter passing only one central

    frequency component, but a band pass filter as-

    sociated with a bank of low pass filters (see Fig. 4).

    The associated low pass filters are beneficial to

    passing the useful low frequency components.

    Therefore, our MGF can more straightforwardly

    express the texture characteristics of fingerprint

    images than the TGF.

    3. Parameter selection

    Parameter selection plays a crucial role in the

    use of the TGF and has long been a research focus

    in the field of image processing. However, the

    computation of filter coefficients is very complex(Bovick et al., 1990). For texture analysis, some

    principles of parameter selection are proposed (e.g.

    Jain and Farrokhnia, 1991; Clausi and Jernigan,

    2000) based on comparison between the output

    of the human visual system and the Gabor filter

    response. Responsible for the specific finger-

    print image enhancement, parameter selection also

    needs to be explored. In the TGF, there are fiveparameters to be specified, including the Gabor

    filter orientation /, the standard deviations rx andry of the 2-D Gaussian function, the period T ofthe assumed sinusoidal plane wave and the con-

    volution mask size 2N 1 2N 1. Honget al. specified them based on the empirical data.

    In our MGF, the period T is decomposed into T1and T2, and most of the parameters including theconvolution mask size are specified adaptively.

    3.1. Orientation / of modified Gabor filter

    Hong et al. (1998) firstly utilized a least mean

    square estimation method to compute the orien-

    tation field of fingerprint images block-wisely. The

    steps are as follows:

    1. Divide the input fingerprint image into blocks

    of size W W .2. Compute the gradients Gx and Gy at each pixel

    x; y in each block.3. Estimate the local orientation of each block us-

    ing the following formula:

    hi;j

    12tan1

    PiW =2uiW =2

    PjW =2vjW =2 2Gxu;vGyu;vPiW =2

    uiW =2PjW =2

    vjW =2 G2xu;vG2yu;v

    !

    9Then, hi; j is regularized into the range of p=2to p=2. Finally, the parameter / of the TGF ischosen as the orientation of each block.

    However, their block-wise scheme is coarse and

    cannot obtain fine orientation field, which tends tocorrupt the TGFs performance. In order to esti-mate the orientation field more accurately, we ex-

    tend their method into a pixel-wise one. For each

    pixel, a block with size W W centered at the pixelis referred to, so the orientation of each pixel can

    be estimated by the formula (9). To reduce the

    computational cost, a sliding window technique is

    employed (Yang et al., 2002). For an image, theorientation of the MGF is tuned to the orientation

    at current pixel, and thus a low pass filtering along

    the orientation and a band pass associated with

    low pass filtering orthogonal to the orientation are

    performed.

    It needs to be emphasized that a step of

    smoothing the orientation field by a low pass filter

    is necessary since sometimes the orientation field isdistorted by noises.

    3.2. Periods T1 and T2

    Examining the formulas (5)(8), we draw the

    conclusion that T1=2 T2=2 in our MGF corre-sponds to T in the TGF that is depicted as theridge distance by Hong et al. (1998) and they be-

    come the same when T1 T2. Further investigatingthe formula (8), we learn that the zero crossings of

    g0 are merely determined by the oscillator F x; T1;T2. Accordingly, we specify T1 and T2 as double

    6 J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx

    ARTICLE IN PRESS

  • of the local ridge width and valley width. The

    determination of T1 and T2 is detailed as follows.Let Ii; j denote an arbitrary pixel to be filtered

    currently whose neighborhoods have the ridge

    width Wr and valley width Wv. Firstly, a method isapplied to roughly determine whether Ii; j is lo-cated on a ridge or valley. If Ii; j belongs to aridge then T1 is set as 2Wr and T2 as 2Wv. Other-wise, if Ii; j is on a valley then T1 is set as 2Wvand T2 as 2Wr (see Fig. 5). The selection of T1and T2 ensures that the centric pixels on eachridges and valleys are given the heaviest weights

    in the later convolution phase, which benefits to

    enhance the contrast between ridges and valleys.To this end, there are two prerequisites to be

    solved:

    1. How to compute the ridge width Wr and valleywidth Wv.

    2. How to determine whether a pixel is on a ridge

    or valley, i.e. segmentation of ridges and val-

    leys.

    Before addressing the two issues, a step of

    smoothing the fingerprint image is necessary be-

    cause it may be filled with noises such as holes on

    ridges and peaks on valleys. We utilize 1-D direc-

    tional Gaussian filter at each pixel along its ori-

    entation to remove the noises.

    3.2.1. Computation of ridge width Wr and valleywidth Wv

    Accurate estimation of the ridge width and

    valley width is in fact a difficult task. We follow

    Hongs method of computing ridge frequency toobtain them. Firstly, the fingerprint image is di-

    vided into blocks of size w w (w 16). For eachblock centered at pixel Ii; j, an oriented windowof size l w (l 32) is built and an x-signaturesignal is computed. Here, the x-signature is theaverage signal of projection of all the intensities in

    the oriented window along the Ii; j orientation(please refer to Hong et al. (1998) for more detailsabout the x-signature). The shape of the x-signa-ture signal is similar to Fig. 1(a) or (b). Its first and

    second order derivatives indicate the ridge width

    and valley width. Inaccurate ridge width and val-

    ley width could lead to inter-block non-uniform

    image enhancement, since the estimation proce-

    dure is block-wise not pixel-wise. To compute

    the ridge width and valley width from the dis-crete signal x-signature more accurately, we resortto a fitting method to acquire the first and second

    order derivatives. Based on the trade-off between

    accuracy and efficiency, the discrete Chebyshev

    polynomials introduced by Haralick (1984) and

    Tico and Kuosmanen (1999) are employed to

    perform the fitting. The zero crossings of the sec-

    ond order derivatives and magnitude of the first

    Fig. 5. The curve of F x; T1; T2 corresponding to different period T1 and T2. Pixel Pa is located on a ridge, so T1 is set as the double ofridge width 2Wr and T2 as the double of valley width 2Wv. Pixel Pb is on a valley, so T1 is set as 2Wv and T2 as 2Wr.

    J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx 7

    ARTICLE IN PRESS

  • order derivatives are together taken into account

    to determine the ridge width and valley width.

    In other words, the distance between two zero

    crossings of the second order derivatives is re-

    garded as the ridge width or valley width if the

    magnitude of the corresponding first order deriv-ative is larger than a threshold. Then, the signs

    of the second order derivatives specify whether it

    is ridge or valley. Thereby, the information of

    ridge width Wr and valley width Wv is associatedto each block. In application, ridge width and

    valley width fall into a certain interval. If ex-

    ceeding the interval, they are replaced by the mean

    of those available in neighboring eight blocks.

    3.2.2. Segmentation of ridges and valleys

    As mentioned above, the functional form of

    F x; T1; T2 depends on the characteristics of cur-rent pixels neighborhoods, and hence differentpixel corresponds to different F x; T1; T2. For thispurpose, a previous step of determining whether

    the pixel is located on a ridge or valley is necessary.In our algorithm, we adopt a local threshold

    method to roughly segment ridges and valleys.

    Firstly, the mean m and standard deviation s ofintensities in each block that is divided in the

    previous phase of estimating ridge width and val-

    ley width are calculated. Secondly, for each block

    a local threshold thres m d s is selected. Fi-nally, each pixel at the block is classified into twocategories of ridge or valley by comparing its in-

    tensity with thres (d 0:2 in our experiments).Generally speaking, this segmentation method

    is rough and some pixels may be misclassified due

    to the existence of noises. But in our experiments,

    the performance is acceptable after Gaussian di-

    rectional smoothing. For more accurate segmen-

    tation, the gradient at each pixel can also beapplied by topography methods (e.g. Wang and

    Pavlidis, 1993; Haralick et al., 1983).

    3.3. Determination of rx and ry

    In the MGF, rx is the standard deviation of the2-D Gaussian function along the x-axis and ryalong the y-axis. rx and ry control the spatialfrequency bandwidth of the MGF response. The

    larger they are, the wider bandwidth is expected.

    However, too wide bandwidth can unexpectedly

    enlarge the noises, and too narrow bandwidth

    tends to suppress some useful signals.

    The value of ry determines the smoothing de-gree along the local orientation. Too large ry canblur the minutiae. In our algorithms, ry is empir-ically set as 4.0.

    Compared with ry , rx inherently plays a moreimportant role for the enhancement performance

    and needs to be specified carefully. It influences the

    degree of contrast enhancement between ridges

    and valleys. This selection involves a trade-off. If

    rx is too large, the factor h0x in formula (5) will havemore high frequency components and even un-

    stably oscillate near the origin, which leads to ar-

    tifacts. On the other hand, if rx is too small, theband pass associated with low pass filters will

    evolve into a pure low pass one due to the over-domination of the Gaussian function in h0xx; T1;T2, which results in blurring edges (boundaries)between ridges and valleys. Hong et al. (1998)empirically selected rx as 4.0 and Greenberg et al.(2000) specified it as 3.0 for his experimental im-

    ages. Both of their parameter selections depend on

    specific image database. It is known that the in-

    fluence of rx on the performance is related with T1and T2 (only T in the TGF). If T1 and T2 are ofgreat variation in a fingerprint image, a constant

    rx could result in a non-uniform enhancement,even in some regions there is no enough enhance-

    ment but in others artifacts occurs.

    The inconsistency of inter-block enhance-

    ment implied that ridges and valleys in differ-

    ent blocks are given non-uniform weights by filter

    masks, since the filtering procedure is a convo-

    lution between images and filter masks. To

    avoid the inconsistency, the MGF mask is as-signed to each block by involving the local char-

    acteristics, T1 and T2. The following constraints areexamined:

    R T1=40

    exp x22r2x

    cos 2pxT1

    dxR T1=4T2=2

    T1=4exp x2

    2r2x

    cos 2pxT1=4T2=4T2

    dx

    Q

    10

    8 J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx

    ARTICLE IN PRESS

  • Z 3T1=4T2=2T1=4T2=2

    exp

    x

    2

    2r2x

    cos 2px T1=2 T2=2T1

    dx 0 11

    Given a fixed Q, rx corresponding to certain T1 andT2 can be obtained by a numeric resolving method.Therefore, constraints (10) and (11) provide a link

    between the rx and T1 and T2, that is, a link be-tween rx and each block. Here, Q represents thearea proportion between the central dominant

    component (near the origin, above the x-axis) andits two close sidelobes (below the x-axis) in thefactor h0x (see Fig. 6). Moreover, constraints (10)and (11) ensure that a MGF is a stable oscillator

    near the origin (Q > 1, in the application), sinceother sidelobes far away from the origin are sup-

    pressed. To achieve a uniform enhancement, Q isspecified as a global one. To speed up the filtering,

    the rxs corresponding to different T1 and T2 arecomputed off-line since the ridge width and valley

    width are in a certain interval. Some rxs adopted inour experiments are listed in Table 1 (Q 1:2).From Table 1, T2 is subdivided into a smaller rangewhen T1 is small.

    3.4. Selection of convolution mask size

    The implementation of enhancing fingerprint

    images by the MGF or TGF is a convolution be-

    tween an image and a part of filters coefficientmatrix. The convolution mask size influences the

    performance of filtering and computational cost.

    Too large mask size tends to burden the en-

    hancement processing and meanwhile bring anunstable factor when the area of the central

    dominant component is less than the sum of that

    of its two close sidelobes. But if it is too small, the

    MGF or TGF collapses into a 2-D low pass filter

    and the advantage of the band pass filter will be

    lost. Hong et al. (1998) set the mask size as

    2N 1 2N 1 (N 5 from his experience.However, it is illogical that the mask size is stillconstant when the width of ridges and valleys

    varies. In contrast, we select the convolution mask

    size as 2Ww 1 2Wh 1 for our MGF whichvaries according to T1 and T2. Here, 2Ww 1 isset as T2=2 T1=2 T2=2 orthogonal to the localorientation (see Fig. 6). Actually, T2=2 T1=2T2=2 means Wv Wr Wv or Wr Wv Wr. Basedon the formulas (10) and (11), this selection en-sures the area of the central dominant is larger

    than the sum of that of its two close sidelobes

    (represented by Q > 1). Thereby, the band passproperty is exerted and meanwhile both instability

    and truncation errors are avoided.

    From the above discussions, the convolution

    mask size is integrated with the T1, T2 and rx by theglobal parameter Q to achieve consistent en-hancement. Moreover, Wh is selected as a constant

    Fig. 6. The response of the factor h0x in formula (8) in spatialdomain. The central dominant component and its close side-

    lobes are marked.

    Table 1

    Some rxs adopted in our experiments corresponding to differentT1 and T2

    T1 T2 rx

    4 4; 12 1.54 14; 18 1.64 20; 28 1.86 1.8

    8 2.5

    10 2.7

    12 3.0

    14 3.5

    16 4.0

    J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx 9

    ARTICLE IN PRESS

  • value 5.0 corresponding to ry specified in theprevious subsection.

    4. Implementation

    In the whole process of image enhancement, the

    MGFs design is completed based on the analysis

    in frequency domain, and images are enhanced in

    spatial domain. Meanwhile, the coefficients of the

    Gaussian directional filter and MGF are com-

    pleted off-line for speedup. In the TGF, Gabor

    filter banks with different orientations are em-

    ployed and their coefficients are computed re-spectively. This entails a number of filters. In our

    Fig. 7. Enhancement results corresponding to the fingerprint images of Fig. 8. The first two columns are the results using the TGF with

    different rx; ry . The third column is the results by our MGF.

    10 J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx

    ARTICLE IN PRESS

  • algorithm, only coefficients of the MGF with the

    orientation / 0 are computed and image rota-tion is implemented instead of computing the

    multi-directional MGFs. That is, MGF banks with

    different rxs corresponding to different T1 and T2are completed in advance. Then, image blocks withthe same size as that of the convolution mask are

    rotated to the MGF orientation / 0. As a result,our MGF enhancement achieves high efficiency,

    although we resort to multi-rx, multi-convolutionmask technique.

    5. Experimental results

    We test the efficiency and robustness of our

    algorithm using some fingerprint images, which

    consist of our image database captured by an op-

    tical live-scanned equipment 400 376,FVC2000 DB2 364 256 (touched sensor), data-base at the University of Bologna 256 256 andNIST 512 512 (National Institute of Standardand Technology) series fingerprint image database.

    The parameters of our MGF are uniform to all the

    images to validate our image-independent param-

    eter selection scheme. Our experimental results

    demonstrate that our MGF is more powerful in

    fingerprint image enhancement than the TGF.

    Some experimental results are illustrated in Fig. 7

    corresponding to the original images in Fig. 8.The experimental results reveal that the difficult

    task in parameter selection of the TGF has been

    resolved in our MGF. The spurious ridges and

    valleys are avoided and uniform enhancement

    performance is achieved. We also performed the

    feature extraction and feature matching (Ratha

    et al., 1996) on a combined fingerprint image

    database from our database, FVC 2000 DB2 and

    the database at University of Bologna. The fin-gerprint matcher reported by Ratha et al. (1996) is

    a widely applied method. It employed the Hough

    transform to align two minutia sets. From the

    experimental results, our MGF makes the feature

    extraction more reliable and feature matching

    more accurate (see Table 2). Further investigating

    our approaches and experiments, we learn that the

    slightly higher computational cost of the MGFprimarily results from its larger convolution mask

    size, since T2=2 T1=2 T2=2 in the MGF is gen-erally larger than the convolution mask width

    2N 1 in the TGF for our tested images (seeTable 3). To achieve fast speed in large images,

    convolution implementations in spatial domain

    can be substituted by the multiplications in fre-

    quency domain.

    Fig. 8. Some fingerprint images in our experiments. (a) is captured from an optical equipment. (b) is f23 of NIST-4. (c) is f09 of NIST-

    4.

    Table 2

    Fingerprint matching performance under the enhanced images

    by the TGF and MGF

    Filter FAR

    FRR

    0.01% 0.05% 0.1% 0.15% 1%

    TGF 5.9% 5.3% 4.3% 3.9% 3.1%

    MGF 3.5% 3.1% 2.9% 2.9% 2.8%

    J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx 11

    ARTICLE IN PRESS

  • 6. Conclusion

    In this paper, a MGF has been proposed for

    fingerprint image enhancement. The modification

    of the TGF can make the MGF more accurate in

    preserving the fingerprint image topography. And

    a new scheme of adaptive parameter selection for

    the MGF is discussed. This scheme leads to theimage-independent advantage in the MGF. Al-

    though there are still some intermedial parameters

    determined by experience, a step of image nor-

    malization can compensate the drawback.

    However, some problems need to be solved in

    the future. A common problem of the MGF and

    TGF is that both fail when image regions are

    contaminated with heavy noises. In that case, theorientation field can hardly be estimated and ac-

    curate computation of ridge width and valley

    width is prohibitively difficult. Therefore, a step of

    segmenting these unrecoverable regions from the

    original image is necessary, which has been ex-

    plored in Hongs work to some extent.

    Acknowledgements

    The authors are highly grateful to the anony-

    mous reviewers for their significant and construc-

    tive critiques and suggestions, which improve the

    paper very much. This work was partially sup-

    ported by Hundred Talents Programs of the Chi-

    nese Academy of Sciences, the Natural ScienceFoundation of China, Grant No. 60172056 and

    697908001, and Watchdata Digital Company. We

    acknowledge that the experiments in this research

    are conducted on the fingerprint database from the

    NIST, University of Bologna and FVC2000. We

    would also like to give thanks to our colleagues in

    National Laboratory of Pattern Recognition for

    their stimulated discussions and comments on our

    work.

    References

    Almansa, A., Lindeberg, T., 2000. Fingerprint enhancement by

    shape adaptation of scale-space operators with automatic

    scale selection. IEEE Trans. Image Process. 9 (12), 2027

    2042.

    Bovick, A.C., Clark, M., Geisler, W.S., 1990. Multichannel

    texture analysis using localized spatial filters. IEEE Trans.

    Pattern Anal. Machine Intell. 12 (1), 5573.

    Clausi, D.A., Jernigan, M.E., 2000. Designing Gabor filters for

    optimal texture separability. Pattern Recognition 33 (11),

    18351849.

    Daugman, J.G., 1980. Two-dimensional spectral analysis of

    cortical receptive field profiles. Vision Research 20, 847856.

    Daugman, J.G., 1985. Uncertainty relation for resolution in

    space, spatial frequency, and orientation optimized by two-

    dimensional visual cortical filters. J. Optical Soc. Amer. 2

    (7), 11601169.

    Gabor, D., 1946. Theory of communication. J. IEE 93, 429

    457.

    Greenberg, S., Aladjem, M., Kogan, D., Dimitrov, I., 2000.

    Fingerprint image enhancement using filtering techniques.

    In: Proc. 15th Internat. Conf. on Pattern Recognition III,

    Barcelona, Spain, pp. 326329.

    Haralick, R.M., 1984. Digital step edges from zero crossing of

    second directional derivatives. IEEE Trans. Pattern Anal.

    Machine Intell. PAMI-6 (1), 5868.

    Haralick, R.M., Watson, L.T., Laffey, T.J., 1983. The topo-

    graphic primal sketch. Internat. J. Robotics Research 2 (1),

    5072.

    Hong, L., Wan, Y., Jain, A.K., 1998. Fingerprint image

    enhancement: Algorithm and performance evaluation.

    IEEE Trans. Pattern Anal. Machine Intell. 20 (8), 777789.

    Jain, A.K., Farrokhnia, F., 1991. Unsupervised texture seg-

    mentation using Gabor filters. Pattern Recognition 24 (12),

    11671186.

    Jiang, X., 2001. A study of fingerprint image filtering. In: Proc.

    8th Internat. Conf. on Image Processing (Biometrics),

    Thessaloniki, Greece, pp. 238241.

    Kamei, T., Mizoguchi, M., 1995. Image filter design for

    fingerprint enhancement. In: Proc. Internat. Symp. on

    Computer Vision, Coral Gables, Finland, pp. 109114.

    Maio, D., Maltoni, D., 1998. Neural network based minutiae

    filtering in fingerprint. In: Proc. 14th Internat. Conf. on

    Pattern Recognition, Brisbane, Australia, pp. 16541658.

    Mehtre, B.M., 1993. Fingerprint image analysis for auto-

    matic identification. Machine Vision, Appl. 22 (6), 124

    139.

    OGorman, L., Nickerson, J.V., 1989. An approach to finger-print filter design. Pattern Recognition 22 (1), 2938.

    Table 3

    Comparison of time cost of fingerprint image enhancement

    (based on P4 1.3 GHz, 128 M RAM PC)

    Image reso-

    lution (pixel)

    TGF (N 5,ry 4:0) rx

    Average time cost (s)

    TGF MGF

    224 288 1.8 0.90 0.92256 256 1.8 0.94 1.01364 256 2.0 1.13 1.22400 376 2.2 1.76 1.89

    12 J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx

    ARTICLE IN PRESS

  • Ratha, N.K., Karu, K., Chen, S., Jain, A.K., 1996. A real-

    time matching system for large fingerprint database. IEEE

    Trans. Pattern Anal. Machine Intell. 18 (8), 799

    813.

    Sherlock, B.G., Monro, D.M., Millard, K., 1994. Finger-

    print enhancement by directional Fourier filtering. IEE

    Proc. Vision Image Signal Process 141 (2), 8794.

    Tico, M., Kuosmanen, P., 1999. A topographic method for

    fingerprint segmentation. In: Proc. 6th Internat. Conf. on

    Image Processing, Kobe, Japan, pp. 3640.

    Wang, L., Pavlidis, T., 1993. Direct gray-scale extraction of

    features for character recognition. IEEE Trans. Pattern

    Anal. Machine Intell. 15 (10), 10531067.

    Willis, A.J., Myers, L., 2000. A cost-effective fingerprint

    recognition system for use with low-quality prints and

    damaged fingerprints. Pattern Recognition 34 (2), 255270.

    Yang, J., Liu, L., Jiang, T., 2002. An improved method for

    extraction of fingerprint features. In: Proc. 2nd Internat.

    Conf. on Image and Graphics, Hefei, PR China. In: SPIE

    4875 (1), pp. 552558.

    J. Yang et al. / Pattern Recognition Letters xxx (2003) xxxxxx 13

    ARTICLE IN PRESS

    A modified Gabor filter design method for fingerprint image enhancementIntroductionTraditional Gabor filter and modified Gabor filterParameter selectionOrientation phi of modified Gabor filterPeriods T1 and T2Computation of ridge width Wr and valley width WvSegmentation of ridges and valleys

    Determination of sigmax and sigmaySelection of convolution mask size

    ImplementationExperimental resultsConclusionAcknowledgementsReferences