EXCLUDING THE REMAINING RIDGES OF FINGERPRINT

IMAGE

En Zhu, Jianping Yin, Chunfeng Hu, Guomin Zhang

School of Computer Science, National University of Defense Technology, Changsha 410073, China

Jianming Zhang

Department of Computer Science, Hunan City University, Yiyang 413049, China

College of Computer and Communication, Hunan University, Changsha 410082, China

Keywords: Fingerprint segmentation, Recoverable, Remaining ridges.

Abstract: Fingerprint segmentation is usually to identify non-ridge regions and unrecoverable low quality ridge

regions and exclude them as background so as to reduce the time of image processing and avoid detecting

false features. In ridge regions, including high quality and low quality, there are often some remaining

ridges which are the afterimage of the previously scanned finger and are expected to be excluded from the

foreground. However, existing segmentation methods do not take the case into consideration, and often, the

remaining ridge regions are falsely taken as foreground. This paper proposes two steps for fingerprint

segmentation aiming to exclude the remaining ridge region from the foreground. The non-ridge regions and

unrecoverable low quality ridge regions are removed as background in the first step, and then the

foreground produced by the first step is further analyzed so as to remove the remaining ridge region. The

proposed method turns out effective in avoiding detecting false ridges and in improving minutiae detection.

1 INTRODUCTION

Fingerprint segmentation is an important problem in

fingerprint recognition. A fingerprint image usually

has to be segmented to remove uninterested regions

before some other steps such as enhancement and

minutiae detection so that the image processing will

consume less CPU time. A fingerprint image

generally consists of different regions: non ridge

region, high quality ridge region, and low quality

ridge region. Fingerprint segmentation is usually to

identify non-ridge regions and unrecoverable low

quality ridge regions and exclude them as

background so as to reduce the time of image

processing and avoid detecting false features and

further to improve the recognition accuracy. Most

segmentation methods are block-wised ones (Mehtre,

1987; Mehtre, 1986; Mehtre, 1989; Mehtre, 1993;

Ratha, 1995; Hong, 1998; Klein, 2002) which divide

the fingerprint image into un-overlapped blocks and

decide on the type (background and foreground) of

each block. And some other methods are pixel-wised

ones (Bazen 2000, Bazen 2001) which determine the

type of each pixel. Fingerprint segmentation

typically computes the feature (or feature vector) of

each element, block or pixel, and then determine the

element’s type based on the feature (vector). The

features used in fingerprint segmentation mainly

include statistical features of pixel intensity,

directional image and ridge projection signal, etc.

Methre (Mehtre, 1987; Mehtre, 1986; Mehtre, 1989;

Mehtre 1993) uses gray variance and the histogram

of pixel gradients in a sub-image block for

segmentation. For each sub-image block Ratha

(Ratha, 1995) computes the variance of the

projection signal on different directions. The

foreground block is of large variance along the

direction orthogonal to the ridges and is of small

variance along the direction parallel to the ridges.

And background is usually of small variance along

all directions. Hong (Hong, 1998) uses the features,

including frequency, variance and the average

difference between the peaks and valleys, of the

ridge projection signal along the direction

orthogonal to the local ridges for segmentation.

Klein (Klein, 2002) computes gray mean, variance,

346

Zhu E., Yin J., Hu C., Zhang G. and Zhang J. (2006).

EXCLUDING THE REMAINING RIDGES OF FINGERPRINT IMAGE.

In Proceedings of the First International Conference on Computer Vision Theory and Applications, pages 346-353

DOI: 10.5220/0001360303460353

 SciTePress

gradient consistency and Gabor response for

segmentation by using HMM. Bazen (Bazen, 2000;

Bazen, 2001) computes gray mean (Bazen, 2001),

variance (Bazen, 2001) and gradient coherence

(Bazen, 2000; Bazen, 2001) to pixel-wisely segment

fingerprint image. Yin (Yin, 2005) also uses

coherence, mean and variance. Jain (Jain, 1997) uses

the output of a set of Gabor filters for segmentation

by adopting clustering. Wang (Wang, 2004) uses

Gaussian-Hermite Moments. Ong (Ong 2003) uses

the orientation coherence for coarse segmentation

and then refine the results by Fourier-based

enhancement, adaptive thresholding, and

postprocessing. And Ren (Ren, 2003) detects feature

dots which are somewhat like ridge edge points to

segment fingerprint image.

Most existing segmentation methods aim to and

are able to exclude regions containing no ridges (e.g.

Fig.1 (d)) or of low quality and hence unrecoverable

(e.g. Fig.1 (c)). Yet none of these methods considers

the excluding of the remaining ridges (e.g. Fig.1

(e)), the afterimage of the previously scanned finger.

And consequently, the remaining ridges are often

falsely taken as the foreground in the case that they

are of clear or recoverable ridge structure. Another

problem in fingerprint segmentation is how to know

whether a low quality ridge block is recoverable or

unrecoverable so as to guide the segmentation. The

typical solution is to visually decide the types,

recoverable and unrecoverable, and feed some

samples, whose type are visually decided, into a

classifier at its training stage, and the trained

classifier would be used to classify the fingerprint

regions. However, in fingerprint image processing,

the process of ridge recovering is done by a certain

algorithm not by manual, and therefore a manually

recoverable ridge block maybe unrecoverable for the

specific algorithm since the algorithm can not be

cleverer than the human brain. Recovering low

quality ridges, e.g. enhancement using a texture

filter by tuning its orientation and frequency (Hong,

1998; Ailisto, 15; Zhu, 2004), usually depends on

the correct computation of ridge orientation.

Incorrect computation of the ridge orientation means

that the ridge can not be recovered. Thus we propose

to segment the fingerprint image through two steps:

The first segments according to the results of ridge

orientation estimation. The recoverable ridge

regions, including high quality and low quality, with

their orientations correctly estimated, are identified

as foreground in the first step. The foreground

identified by the first step may contain remaining

ridge region, and the second step further excludes

the remaining ridge region from the foreground. In

the following sections which will describe the

proposed algorithm in detail, we call the first step

primary segmentation, and the second step

secondary segmentation. Section 2 describes the

primary segmentation. Section 3 describes the

secondary segmentation. Section 4 contains the

experimental results. Section 5 is the conclusion.

And section 6 is the acknowledgement.

(e)

(b)

(c)

(d)

(a)

Figure: 1 Fingerprint regions: (a) high quality ridge

region, (b) recoverable low quality region (the ridge

interrupts are recoverable in this case), (c) unrecoverable

low quality region, (d) non ridge region, (d) remaining

ridge region.

2 PRIMARY SEGMENTATION

Fingerprint segmentation serves for decreasing the

computational expenditure at image processing and

for improving the accuracy of feature (typically

minutiae) extraction, because excluding non-ridge

regions and unrecoverable ridge regions helps to

reduce CPU consumption and avoid introducing

false minutiae, and keeping recoverable ridge

regions not removed helps to avoid losing true

minutiae. However, recoverable ridges are often

actually not recovered in the enhance image,

because they are just manually and not

algorithmically recoverable mainly due to that their

orientations are not correctly estimated. Fig.2 (a) and

(b) show an example of taking algorithmically

unrecoverable ridges, due to the incorrect estimation

of ridge orientation, as foreground. Besides, it is

hard to decide the recoverability of low quality

ridges, and consequently, recoverable ridges, in spite

of the correct estimation of ridge orientations, are

often taken as background. Fig.2 (c) and (d) show an

example of take manually recoverable ridges as

background. Fig.2 (b) and (d) are the segmentation

EXCLUDING THE REMAINING RIDGES OF FINGERPRINT IMAGE

347

(a) (b) (c) (d)

Figure 2: Examples of fingerprint image segmentation by VeriFinger4.2 of Neuro. (a) and (c) are the original images

collected using SecuGen device, and (b) and (d) are the segmented results, respectively.

results by VeriFinger4.2 published by

Neurotechnologija (hereinafter abbreviated as

Neuro) (Neurotechnologija Ltd., 2004) which have

participated in FVC2002 (Maio, 2002) and

FVC2004 (Maio, 2004) and came top in both the

two contests. The main difficulty in fingerprint

segmentation is to answer whether the (low quality)

ridge block is recoverable or unrecoverable by an

automatic algorithm. A well trained classifier may

be able to distinguish high quality ridges and

manually recoverable low quality ridges from other

type of regions. In the case of ridge orientation

estimation following the image segmentation,

although all the foreground blocks are of high

quality or manually recoverable, none of the ridge

orientation estimation algorithms can ensure the

orientation of each foreground block would be

correctly computed, and as a result, those blocks

with their orientations falsely estimated are

practically not recoverable for the recovering

algorithm, such as enhancement (Hong, 1998;

Ailisto, 2003; Zhu, 2004) and ridge tracing (Maio,

1997; Jiang, 2001). Thus, it is reasonable that the

ridge orientation estimation proceeds prior to the

segmentation and that the blocks of falsely estimated

orientation should be taken as background.

The proposed primary segmentation is based on

the work of (Zhu, 2005). (Zhu, 2005) Proposed a

method to estimate the fingerprint image quality by

training a neural network which responds to correct

ridge orientation of ridge block (of high quality or

manually recoverable) with a large value, and

responds with a small value to those blocks which

contain no ridges or contain manually unrecoverable

ridges or are of falsely estimated orientations. For

each image block, a feature vector

1121

,...,, CCC

computed to be fed into the network which will

respond to the vector with a value. The responded

value by the trained network to a specific block is

depended on the orientation, since the items from C5

to C11 of the input vector

1121

,...,, CCC

(Zhu,

2005) have a close relationship with the estimated

ridge orientation. Suppose that the image is divided

into non-overlapped blocks like in (Zhu, 2005), and

let W(i,j) denote the block at the ith row and the jth

column. And the ridge orientation is quantified into

16 orientations: the kth orientation is

/16k

⋅

(0 16)k

≤

. For each block W(i,j), 16 vectors,

denoted as

12 11

, ,...,

CC C

(0 16)k≤<

, can be

computed,

12 11

, ,...,

CC C

corresponding to the

orientation

/16k

⋅

. For each block, feed the 16

vectors to the network and obtain 16 responded

values, respectively. The trained network would

generally respond with large values to the vectors

corresponding to the orientation close to the true

ridge orientation, and respond with small values to

other vectors. Let

[](,)

kij be the responded

value to the kth vector of the block W(i,j). We use

these responded values to each block to estimate the

ridge orientation and primarily segment the image

(primary segmentation) as follows.

∑

−=

+⋅=

),]([)(),](['

jiukRujikR

(1)

∑∑

−=−=

++⋅=

),](['),(),](["

vjuikRvujikR

(2)

(

)

"[ ]( , ) max "[ ]( , ) 0 15R l ij R k ij k=≤≤

(3)

((,)) 16OW i j l

⋅

(4)

1"[](,)

((,))

0"[](,)

lij t

MWi j

lij t

≥

⎧

⎨

⎩

(5)

VISAPP 2006 - IMAGE ANALYSIS

348

(a) (b) (c)

Figure 3: Segmentation failing to remove remaining ridges. (a) original image FVC2004_DB2_35_4, (b) result by Neuro,

where

()u

and

(,)uv

are Gaussian filter and

are used to smooth noisy responded values.

((,))OW i j is the estimated ridge orientation.

((,))

Wij

denotes the result of the primary

segmentation:

(, )Wij is a foreground block if

((,))

Wij =1, and background if

((,))

Wij

=0.

3 SECONDARY SEGMENTATION

The primary segmentation identifies and removes

non-ridge blocks and unrecoverable ridge blocks

(manually unrecoverable or having the falsely

estimated orientations and thus algorithmically

unrecoverable). The foreground of the primary

segmentation contains ridge block of correct

orientation. The remaining ridges of the fingerprint

image tend to be included in foreground, if they

have recoverable clear ridge structure and have their

orientation correctly estimated. It is difficult to

identify remaining ridges by once segmentation,

including the existing segmentation methods and the

proposed primary segmentation as in Fig.3 which

shows the example of segmentation by existing

method and the propose primary segmentation

which fail to remove the remaining ridges, since

remaining ridges often have clear structures and

since it is possible, for two fingerprints A and B as

in Fig.4, the remaining ridges of fingerprint A have

the similar features with or even appear clearer than

the true ridges of fingerprint B. Fortunately, within

the same image, there are typical differences

between the remaining ridges and the true ridges: (1)

the average gray value of the remaining ridge block

is generally bigger than that of the true ridge block;

(2) the difference between ridge and valley in the

remaining ridge block is smaller than in the true

ridge block. The two differences are used by the

secondary segmentation to further identify and

remove the remaining ridges.

(a) (b)

Fig. 4 Remaining ridges and true ridges from two

fingerprints have possible similar features. (a) Fingerprint

A, the left part contains remaining ridges. (b) Fingerprint

B, which contains true ridges which are of similar

features, such as gray value and inter-ridge-valley

difference, with the remaining ridges of Fingerprint A.

Let the LG(W) be the local average gray value of

the block W. The global average gray value of all

the foreground blocks would be

((,))1

(, )

((,))

MWij

Wij

LG W i j

MWi j

∑

(6)

And let LA(W) be the local inter-ridge-valley gray

difference of the block W. The global average inter-

ridge-valley gray difference of all the foreground

blocks is computed as

((,))1

(, )

((,))

MWij

Wij

AW i j

MWi j

∑

(7)

The first difference between the remaining ridge

block and the true ridge block from the same image

EXCLUDING THE REMAINING RIDGES OF FINGERPRINT IMAGE

349

can be described by LG and MG: The value of

GMG− is usually bigger at the remaining ridge

block than at the true ridge block. And the second

difference can be described using LA and MA: The

value of

LA MA−

is usually smaller at the

remaining ridge block than at the true ridge block.

Some blocks which have small LG value and large

LA value can be taken as the true ridge blocks

without regarding to the value of

GMG− and

LA MA−

, and similarly, those blocks which have

large LG value and small LA value can be taken as

the remaining ridge blocks without considering the

value of

GMG− and

AMA− . Therefore,

the secondary segmentation uses 〈LG，MG，LA

，MA〉 to reclassify the foreground blocks of the

primary segmentation. For the blocks from the same

image, they have the same MG value and same MA

value. LMS modal (Press, 1992) is used for the

secondary segmentation. Suppose that N samples,

including positive samples and negative samples, are

selected for training the classifier and are denoted as

{}

NiWyWMAWLAWMGWLG

iiiii

≤≤1)(),(),(),(),(

1)( =

Wy if

W is a positive sample (true ridge

block),

1)( −=

Wy if

W is a negative sample

(remaining ridge block). The LMS modal is

described by equation (8).

⎟

⎠

⎞

⎜

⎝

⎛

⎟

⎠

⎞

⎜

⎝

⎛

⎟

⎠

⎞

⎜

⎝

⎛

)(

...

)(

)()()()(1

...............

)()()()(1

3333

2222

1111

NNNNN

WMAWLAWMGWLG

(8)

where

43210

,,,, aaaaa

are the parameters to be

solved. At the secondary segmentation, given a

block

W , compute

()yW

as equation (9).

01 2

() () ()

() ()

yW a a LGW a MGW

a LAW a MAW

=+⋅ +⋅

+⋅ +⋅

(9)

(( )) 1sign y W =−

where W is a foreground

block in the result of the primary segmentation, take

W as a background block and set ()0MW

Fig.5(c) shows the secondary segmentation result of

image Fig.3(a) which has the remaining ridges not

removed by Neuro and by the primary segmentation.

More results are shown in Section 4.

4 EXPERIMENTAL RESULTS

The experiments use eight images, denoted as

image1~8 respectively: image1, image 2, and

image3 are shown in Fig.2(a), Fig.2(c) and Fig.3(a),

respectively, and images 4~8 are shown in Fig.6.

Fig.4 shows the segmentation results of the first 3

images by the proposed method. And their

segmentation results by the Neuro are shown in

Fig.2(b), Fig.2(d) and Fig.3(b), respectively. Fig.6

shows the segmentation results of the rest 5 images.

Segmentation of fingerprint image serves for

reducing the consumed time of image processing

and improving the accuracy of minutiae detection.

One of method to evaluate an automatic

segmentation method is to compare the segmented

image of the automatic algorithm with the manually

segmented image and then estimate the

segmentation accuracy of the automatic algorithm.

Also, we can use the accuracy of minutiae detection

for comparing two automatic segmentation

algorithms. The accuracy of minutiae detection can

be evaluated using EI (Error Index), EI=(a+b)/t

where a is the number of lost minutiae, b is the

number of spurious minutiae, and t the total number

of minutiae contained in the image. The value of t is

generally computed as the number of manually

labeled minutiae. The smaller the value of EI is, the

more accurate the algorithm is. We quantitatively

verify the proposed segmentation method only using

EI, since the accuracy of segmentation is obviously

shown in Fig.5 and Fig.6. The Error Indexes of

minutiae detection on each experimental image are

listed in Table 1. The average EI of the two

methods, Neuro and the proposed, on the

experimental images are respectively 1.27 and 0.49.

The proposed method produces spurious minutiae

much less than Neuro and greatly decreases the EI

value.

(

)

(

)

(

)

Figure 5: Examples of segmentation by the proposed

method. (a)result of image Fig.2(a), (b)result of image

Fig.2(c), (c)result of image Fig.3(a).

VISAPP 2006 - IMAGE ANALYSIS

350

FVC2002_DB3_48_8FVC2002_DB3_ 95_ 8 FVC2004_DB2_100_6FVC2002_DB3_ 97_ 2 FVC2004_DB2_28_2

Figure 6: Comparison of segmentation between Neuro and the proposed. Left column is the original images; Middle

column is the results by Neuro; Right column is the results by the proposed. The images in left column, from top to bottom,

are respectively denoted as image4, image5, image6, image7 and image8.

EXCLUDING THE REMAINING RIDGES OF FINGERPRINT IMAGE

351

Table 1: EI comparison.

Neuro The proposed Image Manually

labeled

True Spurious Lost EI True Spurious Lost EI

Image1 26 13 25 13 1.46 13 2 13 0.58

Image2 11 3 32 8 3.64 9 4 2 0.55

Image3 29 27 8 2 0.34 26 2 3 0.17

Image4 17 12 10 5 0.88 9 4 8 0.71

Image5 35 29 24 6 0.86 29 2 6 0.23

Image6 17 13 10 4 0.82 10 7 7 0.82

Image7 12 11 12 1 1.08 11 1 1 0.17

Image8 10 6 13 4 1.70 9 6 1 0.70

Average 20 14 17 5 1.27 15 4 5 0.49

4 CONCLUSION

Fingerprint segmentation is not a full-solved

problem in fingerprint recognition and mainly aims

to reduce time expenditure of image processing and

to improve minutiae detection. The main difficulties

of fingerprint segmentation are that low quality

regions are hard to be classified and that the

fingerprint images are often interfered with

remaining ridges which are the afterimage of the

previously scanned finger and are hard to be

removed especially when they appear clear

structures. It is difficult to correctly estimate the

orientations of low quality ridge regions, and a

manually recoverable region should be taken as

background if its orientations are falsely estimated.

Spurious minutiae are generally produced by

including manually or algorithmically unrecoverable

ridge regions as foreground. In order to accurately

remove unrecoverable regions and remaining ridges

and as a consequence to improve the minutiae

detection, this paper, following our previous work

(Zhu, 2005), proposed a method, primary

segmentation, to exclude non-ridge regions and

(manually or algorithmically) unrecoverable regions,

and then proposed the secondary segmentation to

reclassify the foreground blocks of the primary

segmentation to remove the remaining ridges. The

experiments show that the proposed method leads to

an improvement of minutiae detection.

ACKNOWLEDGEMENTS

This work was sponsored by the national natural

science foundation of China (Project No. 60373023)

and by science and technology research Foundation

of Hunan City College (No. 20057306)..

REFERENCES

Ailisto, H., Lindholm, M. (2003). A review of fingerprint

image enhancement methods. International Journal

of Image and Graphics, 3(3), 401-424.

Bazen, A.M., Gerez, S.H. (2000). Directional field

computation for fingerprints based on the principal

component analysis of local gradients. Proc.

ProRISC2000, 11th Annual Workshop on Circuits,

Systems and Signal Processing, Veldhoven, The

Netherlands, November 2000.

Bazen, A.M., Gerez, S.H. (2001). Segmentation of

fingerprint images. Proceedings of Workshop on

Circuits Systems and Signal Processing

(ProRISC2001) pp.276-280.

Klein, S., Bazen, A., & Veldhuis, R. (2002). Fingerprint

image segmentation based on hidden Markov models.

Proceedings of 13th Annual Workshop on Circuits,

Systems, and Signal Processing pp.310-318.

Hong, L., Wang, Y.F., & Jain, A.K. (1998). Fingerprint

image enhancement: algorithm and performance

Evaluation. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 20(8), 777-789.

Jain, A.K., Ratha, N.K. (1997). Object detection using

Gabor filters. Pattern Recognition 30 (2) 295–309.

Jiang, X., Yau, W.Y., & Ser, W. (2001). Detecting the

fingerprint minutiae by adaptive tracing the gray-

level ridge. Pattern Recognition, 34, 999-1013.

Maio, D., Maltoni, D. (1997). Direct gray-scale minutiae

detection in finger-prints. IEEE Transactions Pattern

Analysis and Machine Intelligence, 19(1), 27-39.

Maio, D., Maltoni, D., Cappelli, R., Wayman, J.L., &

Jain, A.K. (2002). FVC2002: second fingerprint

verification competition. Proceedings of 16th

International Conference on Pattern Recognition

(ICPR'02) Volume 3 August 11 - 15. (http://bias.csr.

unibo. it/fvc2002.)

Maio, D., Maltoni, D., Cappelli, R., Wayman, J.L., &

Jain, A.K. (2004). FVC2004: third fingerprint

verification competition. Proceedings of

International Conference on Biometric

Authentication, (2004). (http://bias.csr.unibo.it/

fvc2004)

VISAPP 2006 - IMAGE ANALYSIS

352

Mehtre, B.M., Murthy, N.N. (1986). A minutiae-based

fingerprint identification system, 2nd Int. Conf.

Advances in Pattern Recognition and Digital

techniques, Calcutta.

Mehtre, B.M., Murthy, N.N., & Kapoor, S. (1987).

Segmentation of fingerprint images using the

directional image. Pattern Recognition, 20(4), 429-

435.

Mehtre, B.M. (1989). Segmentation of fingerprint images

– a composite method. Pattern Recognition, 22(4),

381-385.

Mehtre, B.M. (1993). Fingerprint image analysis for

automatic identification. Machine Vision and

Applications, 6, 124-139.

Neurotechnologija Ltd., (2004). Verifinger 4.2,

http://www.neurotechnologija.com/download.html.

Ong, T.S., Andrew, T.B.J., David, N.C.L., & Sek, Y.W.

(2003). Fingerprint images segmentation using two

stages coarse to fine discrimination technique.

Proceedings of 16th Australian Joint Conference on

Artificial Intelligence, Perth Australia, LNAI 2903,

pp. 624-633.

Press, W.H., Flannery, B.P., Teukolsky, S.A., &

Veherling, W.T. (1992). Numerical Recipes in

C: The Art of Scientific Computing, Cambridge

University Press, 1992.

Ratha, N., Chen, S., & Jain, A.K. (1995). Adaptive flow

orientation-based feature extraction in fingerprint

images. Pattern Recognition, 28(11), 1657-1672.

Ren, Q., Tian, J., & Zhang, X.P. (2002). Automatic

segmentation of fingerprint images. Proceedings of

the Third Workshop on Automatic Identification

Advanced Technologies (AutoID 2002), Oral Report,

Tarrytown, New York, USA, pp.137-141.

Wang, L., Dai, M., & Geng, G.H. (2004). Fingerprint

image segmentation by energy of Gaussian-Hermite

moments. Proceedings of Sinobiometrics 2004,

LNCS 3338, pp.414–423.

Yin, Y.L., Wang, Y.R., & Yang, X.K. (2005). Fingerprint

image segmentation based on quadric surface model.

Proceedings of Audio- and Video-based Biometric

Person Authentication, LNCS 3546, pp.647-655.

Zhu, E., Yin, J.P., & Zhang, G.M. (2004). Fingerprint

enhancement using circular Gabor filter. Proceedings

of International Conference on Image Analysis and

Recognition, LNCS 3212, pp. 750-758.

Zhu, E., Yin, J.P., Hu, C.F., & Zhang，G.M. (2005).

Quality estimation of fingerprint image based on

neural network. Proceedings of International

Conference on Natural Computing, LNCS 3611,

pp.65-70.

EXCLUDING THE REMAINING RIDGES OF FINGERPRINT IMAGE

353