Fingerprint Recognition Based on Combined Features

Yangyang Zhang, Xin Yang, Qi Su, and Jie Tian

Center for Biometrics and Security Research, Key Laboratory of Complex Systems
and Intelligence Science, Institute of Automation, Chinese Academy of Sciences,
Graduate School of the Chinese Academy of Sciences,
P.O.Box 2728 Beijing 100080 China
[email protected],[email protected]
http://www.fingerpass.net

Abstract. In this paper, we represent the fingerprint with a novel local feature de-
scriptor, which is composed of minutia, the sample points on associated ridge and
the adjacent orientation distribution. Then a novel fingerprint recognition method
is proposed combining the orientation field and the local feature descriptor. We
compare two descriptor lists from the input and template fingerprints to calcu-
late a set of transformation vectors for fingerprint alignment. The similarity score
is evaluated by fusing the orientation field and the local feature descriptor. The
experiments have been conducted on three large-scale databases. The compari-
son results approve that our algorithm is more accurate and robust than previous
methods based on the minutiae or ridge features, especially for those poor-quality
and partial fingerprints.

Keywords: orientation field, local feature descriptor, fingerprint alignment, fus-

ing, similarity score.

1 Introduction

Fingerprints are ridge patterns flowing on the surface of fingers, which have been
increasingly used for individual identification in the civilian’s daily life due to the
uniqueness, permanence and universality. Previously researchers have proposed many
kinds of fingerprint recognition algorithms. The most popular methods are based on
point features (called minutiae), which extract minutiae sets from the input fingerprint
and compare them with those from the template fingerprint to calculate the similar-
ity score [1] [2]. The minutiae-based features are simple for storage and effective for
comparison, but there are still some limits in fingerprint recognition. First, the accuracy
of minutiae extraction relies on local quality of fingerprint. Second, minutiae cannot
characterize the overall pattern of a fingerprint, and it is hard to further improve the
performance. So researchers proposed other representations of fingerprints to resolve
these problems, such as texture, ridge structure, orientation field and so on.
The ridge feature observes more information in the whole region of fingerprint and
reinforces the individuality of fingerprint considerably. Tian et. al [3] constructed the
minutia-simplex which contains a pair of minutiae with the associated local ridge in-
formation to estimate the parameters of affine transform and calculate the similarity

S.-W. Lee and S.Z. Li (Eds.): ICB 2007, LNCS 4642, pp. 281–289, 2007.
c Springer-Verlag Berlin Heidelberg 2007
282 Y. Zhang et al.

score. Feng et. al [10] established both the ridge correspondences and the minutia cor-
respondences between two fingerprints. The method realized fingerprint verification
after alignment by incrementally matching ridges and minutiae. Both of the algorithms
survive the problem of the alignment inaccuracy, which influences their performance in
some extent. Another kind of important features is orientation field in fingerprint pro-
cessing. It describes the global structure of the fingerprint ridge pattern. Gu et. al [5]
represented and verified the fingerprints combing both minutiae and model-based ori-
entation field. The algorithm performs well on FVC2002 databases, but in reality it’s
difficult to estimate the global model for some partial fingerprints.
In this paper we represent the fingerprint with a novel local feature descriptor, which
is composed of minutia, the sample points on associated ridge and the adjacent orien-
tation distribution. A novel approach is introduced for fingerprint recognition by fusing
the orientation field and the local feature descriptor. First, we align two fingerprints
by the local feature descriptor and compute the transformation parameters including
the translation and rotation. Second, we combine the local feature descriptor and ori-
entation field feature in fingerprint matching. The comparison of orientation field can
reduce the false accepted rate to accelerate the matching process. The similarity score is
calculated based on combined features with the SUM rule, which is the most accurate
in the experiments. We conduct a set of experiments on three large-scale databases and
make comparisons with the state-of-the-arts. The results show that the fusion of local
feature descriptor and orientation field can certainly obtain better fingerprint authenti-
cation performance than previous methods based on minutiae or ridge features.
The rest of the paper is organized as follows: Section 2 proposes the fingerprint
representation with the local feature descriptor. Section 3 presents the method of align-
ment by using the local feature descriptor. Section 4 describes the process of fingerprint
matching based on both the orientation field and the local feature descriptors. The ex-
perimental results for our algorithm are displayed in section 5 and Section 6 summarizes
our researches.

2 Fingerprint Representation

We represent the fingerprint with a set of local feature descriptors, each of which con-
sists of a minutia, the sample points on its associated ridge and the adjacent orientation
distribution. The minutiae set denoted by {mk = (xkm , ym k k
, θm )|1 ≤ k ≤ K} are ex-
tracted from the thinned fingerprint image with the conventional methods, where mk is
the kth minutia, K is the number of minutiae, (xkm , ym k
) and θmk
∈ [0, π) are respec-
k
tively the location and orientation for m . In this section, we mainly discuss how to
construct the local feature descriptor for the minutiae of different types.
According to the traditional methods, minutiae are classified into two types: ridge
termination and bifurcation. In practical it is rather difficult to distinguish them because
of the noise and environment during fingerprint capture. The ridge ending is possibly
mistaken as the bifurcation due to ridge conglutination, while the bifurcation may turn
into the ending due to the broken ridge.
To solve this problem, we sample both the associated ridge and valley in a constant
interval to express the ridge features in a common form. The ridge feature structure is
 Fingerprint Recognition Based on Combined Features 283

constructed with the information of sample points assigned to minutia as shown in Fig. 1.
We define the set of ridge structures R = {rk }K
k−1 in a fingerprint as follows:

Lk
rk = {{xkt,i , yt,i
k
, lenkt,i , Δφkt,i }t=0,1 }i=1
t
. (1)

Δφkt,i = d(θt,i
k k
, θm ) (2)
⎧
⎨ θ1 − θ2 if abs(θ1 − θ2 ) ≤ π/2
d(θ1 , θ2 ) = θ1 − θ2 − π if (θ1 − θ2 ) > π/2 (3)
⎩
θ1 − θ2 + π if (θ1 − θ2 ) < −π/2
where rk is the local ridge structure belonging to the kth minutia. t ∈ {0, 1} denotes
the sampling type. When the sample points on the ridge for bifurcation or on the valley
for ending, the value of t is set 0, otherwise, it’s set 1. Lkt denotes the number of sample
points. (xkt,i , yt,i
k
) are the coordinates of the sample points. lenkt,i , θt,i
k
are the distance
k
and orientation from the kth minutia m to the ith sample point and θt,i k
∈ [0, π).
d(θ1 , θ2 ) is called the orientation distance taking the value in [−π/2, π/2].The ridge and
valley is traced until encountering another minutia or arriving the border of fingerprint.
If the sampling number Lkt is less than lT hr (In our experiments, lT hr = 3), the
associated ridge structure is considered false and discarded.

(a) (b)

Fig. 1. (a) sampling ridges in two directions for two types of minutiae, (b) ridge structure in one
direction for the minutia m

We also characterize each minutia with a structure that comprises information for
the orientation field in a broad region around the minutia point. The structure comprises
the sampling points equally distributed in a matrix pattern centered at each minutia
as illustrated in Fig. 2. According to the sampling principle, the sampling interval v
is selected as 2τ (τ is the average fingerprint ridge width) to obtain the best tradeoff
between decreasing storage and preserving information.
We define the set of orientation structures V = {vk }Kk=1 in the fingerprint as follows:

vk = {(Δϕki,j )N
j=−Nk }i=−Nk
k Nk
(4)
284 Y. Zhang et al.

Δϕki,j = d(θi,j
k k
, θm ) (5)
where K is the number of minutiae, 2Nk + 1(1 ≤ k ≤ K) is the dimension of sampling
matrix around the kth minutia and each sampling matrix consist of (2Nk +1)∗(2Nk +1)
points pki,j (−Nk ≤ i, j ≤ Nk ). θi,j
k
∈ [0, π) denote the orientation at each sample point.

Fig. 2. The orientation distribution around a minutia (the red point), the circle denote the sample
points Pi,j (Nk = 3, v=16)

The local feature descriptor is proposed for each minutia including its location and
orientation, the associated ridge structure and the adjacent orientation structure. We
represent the fingerprint with a set of local feature descriptors as follows:

M = {M k |1 ≤ k ≤ K} and M k = {mk , rk , v k } (6)

3 Fingerprint Alignment
The alignment stage determines the correct transformation between two fingerprints,
which plays a crucial role in fingerprint recognition. In our approach, we realize the
alignment by finding N most similar local feature descriptor pairs and calculating the
corresponding transformation vectors as candidates. Compared with using the best one
for alignment, the muti-candidates method is more robust to those incorrect and dis-
placed descriptors.
A local similarity function between two descriptors M f = {mf , rf , v f } and M t =
{mf , rf , v f } is defined as:

Sm(M f , M t ) = Sr(rf , rt ) ∗ Sv(v f , v t ) (7)

where Sr(r , r ) and Sv(v , v ) denote the similarity of ridge and orientation structure.
f t f t

In the local descriptor, lenkt,i , Δφkt,i and Δϕkt,i describe the transformation-invariant
relative features of M k , while (xkm , ymk k
, θm ), θt,i
k k
and θi,j represent the transformation-
variant relative features. We utilize the transform-invariant parts of the local feature
 Fingerprint Recognition Based on Combined Features 285

descriptor for their similarity calculation. The similarity functions of ridge and orienta-
tion structure are given in an exponential form [7]:

1  1
Sr(rf , rt ) = exp{− (α ∗ |lenfp − lentp | + (1 − α) ∗ |Δφfp − Δφtp |)} (8)
Np μ
p∈P

1  1
Sv(v f , v t ) = exp{− ∗ |Δϕfq − Δϕtq |} (9)
Nq ω
q∈Q

where P = {(t, i)} denote the overlapped part of the sample points on the ridges.
Q = {(i, j)} indicate the overlapped part of the sample points of orientation field. Np
and Nq are the size of P and Q. α ∈ [0, 1] adjusts the scale difference between the
distance and orientation. The values of μ, ω are empirically set in training.
f
ti K t
Let F = {M f i }K f i=1 and T = {M }ti=1 denote the descriptor sets detected from
the input and template fingerprints. The similarity score is calculated between any pair
of local feature descriptors M f i and M ti . A relative array C is established to store these
index pairs (f i, ti) that satisfy Sm(M f t , M ti ) > SmT hr (SmT hr is an experimental
variant) in the decrease order of their similarity values.
We construct the alignment candidate set L to store the transformation vectors Ln
between these descriptor pairs, which have the largest N similarity score L = {ln }N n=1 .
For any pair of M f i and M ti in the candidate set, we average the difference of the
position and orientation between two minutiae and the corresponding sample point pairs
on the ridges to calculate their transformation vector which includes the translation and
rotation parameters, ln = (dx, dy, dθ)T :

1  t
dθ = θp − θpf (10)
Np
p∈P

     f
dx 1  xtp x
= ( t − Rdθ ∗ fp ) (11)
dy Np y p yp
p∈P

Where Rdθ is a 2 ∗ 2 operator of clockwise rotating with dθ. P and Np are defined in
formula 8.

4 Fingerprint Matching

4.1 Matching Based on Orientation Field

In our approach the block orientation field is utilized to avoid misaligning two finger-
prints, it measures the similarity of their global ridge patterns. For each transformation
vector in the candidate set, we align the orientation fields extracted from the enhanced
fingerprints, where the backgrounds have been excluded. The overlapped area A is par-
titioned into a lattice of blocks of size n ∗ n (n is decided by ridge frequency and
fingerprint quality).
286 Y. Zhang et al.

For each block B, let gs = (gsx , gsy ) denotes the gradient intensity at each site s. We
evaluate the local quality of the fingerprint by using the covariance matrix J and the
normalized coherence measure  k ∈ [0, 1] in the block B as follows [8]:
 
1  j11 j12
J= 2 gs gs =
T
(12)
n j21 j22
s∈B

 j 2 − j22
2
+ 4j122
k = 11 2
(13)
(j11 + j22 )
The similarity score of two orientation fields can be computed as the weighted aver-
age of the block-wise similarity measures:
1 
So(Of , Ot ) = S(B f , B t ) ∗ f
kB ∗t
kB (14)
 ∗
k f
k t
B∈A B B B∈A

1 f t
S(B f , B t ) = exp(− ∗ |d(θ B , θB )|) (15)
v
f t f t
where θB , θB are the average orientation in the block B of two fingerprints, d(θ B , θB )
denotes the orientation distance as defined in the formula 3. v is experimentally selected
based on the training results.
If So(Of , Ot ) < SoT hr, the corresponding local descriptor pair is considered mis-
aligned and discarded from the candidate set, which reduces the false accepted rate and
accelerates the matching process. SoT hr is selected according to false reject rate (in
our method at F RR = 0.1%) in training.

4.2 Matching Based on Local Feature Descriptors

After refining the alignment candidates with the block orientation field, we apply each
residual transformation vectors V n in the candidate set V to the two descriptor sets F
and T from the input and template fingerprints. The similarity score of fingerprints is
evaluated based on local descriptor as the following stages:
– Initialize the set DP to store the corresponding matched descriptor pairs.
– Convert the descriptor set T in the template fingerprint into the aligned descriptor
set T  with the transformation vector.
– Represent all the minutiae from the two sets F and T  in a polar coordinate system
whose origin is set at the minutia in M f i . Because of the non-linear deformation
and the error in feature extracting, the corresponding minutiae from two finger-
prints can’t be exactly overlapped. If two minutiae fall in the same changeable
tolerance box [8], which means, their Euclidean distance and orientation difference
don’t exceed the value δd and δθ, two corresponding descriptors are considered
as the matched pair. We delete those descriptor pairs that don’t locate in the same
tolerance box from the relative array C mentioned in section 3.
– Construct the set of matched descriptor pairs DP with the dynamic programming
method. The descriptor pair (M f m , M tm ) that has the largest similarity value is
inserted to the set DP . All the pairs comprising the two descriptors are discarded
from the array C. This step is repeated until the array is empty.
 Fingerprint Recognition Based on Combined Features 287

– Evaluate the global similarity score between all the matched descriptor pairs. Con-
sidering only the descriptors in the overlapped area after alignment, the similarity
is computed as follows:

1 
M
( i=1 N s[i])2
M
Sp(F, T ) = ∗ Sm(M f m[i] , Sm(M tm[i] ) ∗ (16)
M i=1 N sf ∗ N st

where M is the number of the matched descriptor pairs (f m[i], tm[i]). N s[i] is the
number of the matched sample points in the ith descriptor pair. N sf and N st respec-
tively denote the number of all the sample points from the two descriptor sets F and T 
in the overlapped area after fingerprint alignment.

4.3 Evaluating Combined Similarity Score

Fusing different features in matching stage can improve the fingerprint authentication
performance. The detailed analysis on classifier fusion can be seen in the literatures
[9] [10]. We combine the orientation field and local feature descriptors by different
strategies in score level. The combined similarity score is evaluated in our approach
using four different decision-level fusion rules as follows:
Max rule: Score = max(Sp, So)
Min rule: Score = min(Sp, So)
Product rule: Score = Sp ∗ So
Sum rule: Score = q ∗ Sp + (1 − q) ∗ So, where q ∈ [0, 1] adjusts the weight fractions
between the similarity of orientation fields and local descriptors, it is selected based on
the quality of fingerprints.

Table 1. Comparison of the performance on FVC2004 DB1

FVC2004 DB1 Only orientation-based Only descriptors-based Max Min Product Sum
EER (%) 6.53 4.74 4.56 4.95 4.35 3.49
FMR100 (%) 17.33 11.97 10.42 11.72 9.71 6.21
FMR1000(%) 27.61 23.50 23.29 19.47 15.36 19.04
ZeroFMR(%) 38.43 28.97 28.97 23.45 19.93 24.61

We evaluate the set of combined similarity score according to each transformation

vector in candidates. The maximum value is selected as the final matching score. Ex-
perimental results performed on DB1 of FVC2004 with four fusion rules are displayed
in Table 1. It is obvious that the performance of the Sum rule is the best, while the per-
formance of the Min rule is the worst. Hence, the Sum rule is selected in our algorithm
to evaluate the combined similarity score between the template and input fingerprints.

5 Experimental Results
The experiments have been conducted on FVC2004 [11] and one database of our lab.
Each database of FVC2004 consists of 800 fingerprint images (100 fingers 8
288 Y. Zhang et al.

impressions per finger) captured at a resolution of about 500dpi. Databases empha-

size on distortion, dry and wet fingerprints especially in DB1 and DB3 of FVC2004,
which are more difficult than FVC2000/FVC2002. Therefore our algorithm is evaluated
on DB1 and DB3 of FVC2004. We also establish the database to test the efficiency of
our algorithm for poor-quality and partial fingerprints. The database comprises 1680
fingerprints (140 fingers 12 impressions per finger) acquired with an electric field sen-
sor ”Authentec 4000”. The image size is 96*96 pixels at a resolution of 250 dpi. Due
to the low resolution and small size of the sensor, the fingerprint ridges are blurry and
the impressions from the same finger usually have small overlapped area. The finger-
prints are collected from heterogeneous volunteers including manual workers and el-
derly people, so the majority of them suffer the influence from blur, scar, wetness or
dryness.
The experiments are conducted on three databases (DB1, DB3 of FVC2004 and
our established database). All the matching performance in this section is evaluated
according to the experimental protocols in [12]. To approve the efficacy of the combined
features, we display the performance of the ridge-based method [5] and the minutiae-
based method [1] for comparison in Table 2. The results confirm that the combined
features can significantly improve the performance of fingerprint matching, especially
for the poor-quality and partial fingerprints.

Table 2. Comparison of performance between our algorithm and other matching methods

EER(%) FMR100(%) FMR1000(%) ZeroFMR(%)

Minutia-based 5.74 9.35 15.54 18.43
DB1 Ridge-based 4.57 6.71 16.10 31.18
Our algorithm 3.49 6.21 19.04 24.61
Minutia-based 2.78 4.47 8.40 11.86
DB3 Ridge-based 2.11 3.04 8.75 16.86
Our algorithm 1.31 1.61 4.29 9.15
Our Minutia-based 6.55 10.89 17.54 39.18
established Ridge-based 5.01 5.51 8.28 30.69
database Our algorithm 3.33 4.47 19.39 33.54

6 Conclusion

In this paper we represent the fingerprint with a novel local feature descriptor, which
is comprehensive and robustly holds the uniqueness of the fingerprint. We align the
fingerprints using the descriptors and calculate a set of transform vectors as candidates.
In matching process, the orientation field is used to avoid misaligning and to measure
the similarity score of global ridge pattern. We also evaluate the similarity between
the local descriptor sets from the input and template fingerprints. The two methods
are fused with the Sum rule to obtain the final matching score. The experimental results
demonstrate that our algorithm integrates the advantages of the local and global features
and improves the algorithm performance and reliability.
 Fingerprint Recognition Based on Combined Features 289

Acknowledgments. This paper is supported by the Project of National Science Fund for
Distinguished Young Scholars of China (No. 60225008), the Key Project of National Nat-
ural Science Foundation of China (No. 60332010 and 60575007), the Project for Young
Scientists’ Fund of National Natural Science Foundation of China (No.60303022), and
the Project of Natural Science Foundation of Beijing (No.4052026).

References
1. Jain, A.K., Hong, L., Bolle, R.: On-Line Fingerprint Verification. IEEE Trans. Pattern Anal-
ysis and Machine Intelligence 19(4), 302–314 (1997)
2. Kovcs-Vajna, Z.M.: A Fingerprint Verification System Based on Triangular Matching and
Dynamic Time Warping. IEEE Trans. Pattern Analysis and Machine Intelligence 22(11),
1266–1276 (2000)
3. Tian, J., He, Y.L., Chen, H., Yang, X.: A Fingerprint Identification Algorithm by Clustering
Similarity. Science in China Ser. F Information Sciences 48(4), 437–451 (2005)
4. Feng, J.J., Ouyang, Z.Y., Cai, A.: Fingerprint Matching Using Ridges. Pattern Recogni-
tion 39, 2131–2140 (2006)
5. Gu, J.W., Zhou, J., Yang, C.Y.: Fingerprint Recognition by Combining Global Structure and
Local Cues. IEEE Transaction on Image Processing 15(7), 1952–1964 (2006)
6. Tico, M., Kuosmanen, P.: Fingerprint Matching Using an Orientation-based Minutia Descrip-
tor. IEEE Trans. on Pattern Analysis and Machine Intelligence 25(8), 1009–1014 (2003)
7. Chen, Y., Dass, S.C., Jain, A.K.: Fingerprint Quality Indices for Predicting Authentica-
tion Performance. In: Proc. of Audio- and Video-based Biometric Person Authentication
(AVBPA), pp. 160–170. Rye Brook, NewYork (2005)
8. Luo, X.P., Tian, J.: Image Enhancement and Minutia Matching Algorithm in Automated
Fingerprint Identification System. Journal of software 13 (2002)
9. Kittler, J., Hatef, M., Duin, P.W., Matas, J.: On combining classifiers. IEEE Trans. Pattern
Analysis and Machine Intelligence 20(3), 226–239 (1998)
10. Ho, T.K., Hull, J.J., Srihari, S.N.: Decision combination in multiple classifier systems. IEEE
Trans. Pattern Analysis and Machine Intelligence 16(1), 66–75 (1994)
11. http://bias.csr.unibo.it/fvc2004/
12. Maio, D., Maltoni, D., Cappelli, R., Wayman, J.L., Jain, A.K.: FVC2004: Third Fingerprint
Verification Competition. In: Zhang, D., Jain, A.K. (eds.) ICBA 2004. LNCS, vol. 3072, pp.
1–7. Springer, Heidelberg (2004)