1358
10, OCTOBER 2004
Markov Random Field Models for Directional Field
and Singularity Extraction in Fingerprint Images
Sarat C. Dass
Abstract—A Bayesian formulation is proposed for reliable and
robust extraction of the directional field in fingerprint images using
a class of spatially smooth priors. The spatial smoothness allows
for robust directional field estimation in the presence of moderate
noise levels. Parametric template models are suggested as candi-
date singularity models for singularity detection. The parametric
models enable joint extraction of the directional field and the singu-
larities in fingerprint impressions by dynamic updating of feature
information. This allows for the detection of singularities that may
have previously been missed, as well as better aligning the direc-
tional field around detected singularities. A criteria is presented
for selecting an optimal block size to reduce the number of spu-
rious singularity detections. The best rates of spurious detection
and missed singularities given by the algorithm are 4.9% and 7.1%,
respectively, based on the NIST 4 database.
Index Terms—Bayesian statistics, directional field estimation,
Markov random field models, singularity detection.
I. INTRODUCTION
F IGURE 1 shows a fingerprint impression, of size
512 512, consisting of smoothly varying flow-like
patterns (termed as ridge structures), together with important
singularities (termed as cores and deltas). The direction of
flow of the ridge structures at each location in the image can
be represented as a two-dimensional orientation vector with
unit norm. The directional field is defined as the collection of
orientation vectors for all sites in the image. Singularities (cores
and deltas) are points of discontinuity of the flow field. The two
types of singularities are defined in terms of the ridge structures
[1]; the core is the end point of the innermost curving ridge
while the delta is the confluence point of three different flow
directions (see Fig. 1). The directional field and singularities
represent two fundamental features of fingerprint impressions
that need to be extracted reliably for subsequent processing [2].
Obtaining fast and reliable estimates of the directional field
has been the focus of many previous research efforts; they in-
clude methods based on neural networks [3], filter-based ap-
proaches [4], and gradient-based approaches [2], [5]–[8]. How-
ever, none of these methods explicitly model the inherent spatial
smoothness in the ridge structures. The detection of singulari-
ties has been addressed in many previous work. Finding regions
of high curvature and subsequently classifying a feature vector
into either core, delta or none of these is the approach taken
Manuscript received February 23, 2003; revised January 16, 2004. This work
was supported by the NSF ITR under Grant 0312646. The associate editor co-
ordinating the review of this manuscript and approving it for publication was
Dr. Nicolas Rougon.
The author is with the Department of Statistics and Probability, Michigan
State University, East Lansing, MI 48824 USA (e-mail: sdass@msu.edu).
Digital Object Identifier 10.1109/TIP.2004.834659
1057-7149/04$20.00 © 2004 IEEE
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 13, NO.
Fig. 1. Ridge structures and singularities in a typical fingerprint impression.
in [9], [10]. In [11], geometric theory of differential equations
is used to derive signal-to-symbol representations in the flow
field domain for cores and deltas. The local energy of the direc-
tional field in a neighborhood of a site is used to measure how
closely it resembles a flow field around singularities in [12]. In
[13], a ratio of two sines of directional fields in two adjacent
regions is used to detect singularities while [2] uses a scheme
for detecting singularities based on the Poincare index of the
squared directional field. The present work addresses the esti-
mation of the directional field and the detection of singulari-
ties in a Bayesian framework. A class of spatially smooth sta-
tistical models, having a Markov random field structure, is pro-
posed as priors for the directional field. Markov random field
models have been used with great success for the solution of
a number of important image processing problems in which
regularization based on spatial proximity is critical; these in-
clude applications in image restoration [14]–[16], segmentation
[17]–[20], boundary detection [21], [22], and reconstruction in
inverse problems [23], [24]. See also [25] and [26], and refer-
ences therein. Other regularization techniques have also been
used for image restoration (see [27] and [28], for example). The
spatial smoothness induced through the Markov random field
prior models restrict random variations (for example, caused by
assignable noise factors such as fingertip pressure, skin elas-
ticity, and random smudges during fingerprint sensing) of direc-
tional field vectors that are spatially close, resulting in a robust
extraction algorithm. Based on the robust directional field esti-
mate, we also propose detecting singularities using parametric
template models. One advantage of the parametric modeling is
that a fewer number of spurious detections will be made in pres-
ence of moderate noise levels due to the restrictions placed on
the directional field pattern around singularities.
In all previous work, the detection of singularities assumed
that a reliable directional field had already been extracted. One
disadvantage of this approach is that the detection of a singu-
larity does not have any impact for subsequent molding of the
DASS: MARKOV RANDOM FIELD MODELS 1359

directional field; thus, information about singularities is not used measures the similarity between the two orientation vectors with
for updating the values of the orientation vectors. We show that . The maximum (minimum) value ( ) is
dynamic updating of the directional field using information on achieved when ( ). Given , the distribution of
extracted singularities help detect other singularities that may is given by the density
have been missed previously. Our approach here will be to ex-
tract both the directional field and singularities simultaneously (3)
in fingerprint impressions. Thus, information on the flow field where is the precision constant, and is the ap-
and singularities are utilized simultaneously for feature extrac- propriate normalization (independent of ) so that the density
tion resulting in improved performance. We also investigate the in (3) integrates to unity. Thus, the density in (3) is a monotonic
appropriateness of different block sizes for extracting the flow function of the similarity measure with modes at .
field and singularities. Assuming independence for the different sites in , the ex-
We give an overview of the sections in the rest of the paper. pression
Section II presents a class of Markov random field models for
the extraction of the directional field in fingerprint images. Sec- (4)
tion III discusses the detection of singularities in fingerprint im-
ages based on a similarity measure that determines the closeness represents the joint density of given . Equation (4) also rep-
of the observed directional field to template models for cores resents the likelihood of in terms of the observed data.
and deltas. Section IV presents the joint extraction of the di- It is common in fingerprint analysis to compute a single value
rectional field and singularities by combining the methodology of the directional field for blocks of sites of size , say, by
outlined in Sections II and III. Section V presents the results of grouping the sites in . Blocking results in a spatial lattice ar-
the joint feature extraction on fingerprint images in the NIST rangement of blocks, ,
database. It is shown that the quality of extraction is a function of size . We denote the collection of block indices in the
of the blocking size, and so a criteria is presented for selecting reduced lattice arrangement,
a block size that optimizes the quality of the extraction. by .Blocking assumes that the principal gradient direc-
tions are constant in every block of pixels. Thus, for ,
II. ROBUST MODELS FOR DIRECTIONAL FIELD EXTRACTION we assume that for all pixels . Under blocking, we
Several difficulties arise during the estimation of the direc- have and repre-
tional field: 1) the gradient of the image intensity fluctuates in senting the collection of block-wise principal gradient directions
sign due to the presence of alternating ridges and valleys, al- and directional field values, respectively. For a site , the
though the flow direction remains the same, and 2) it is not pos- precision constant is represented as where
sible to distinguish between opposite flow directions (flow at is a measure of overall precision, measures the contribu-
angle to the horizontal axis is the same as the flow along angle tion of block to the overall likelihood, and is the weight
). Our aim in this section is to propose models that cap- assigned to site within block . With this parameterization of
ture the inherent spatial smoothness of the directional field while , the likelihood in (4) reduces to
taking 1) and 2) into account.
Consider a lattice domain consisting of rows and
columns. An image of size will be represented by its
gray intensity values for each point . Denote
by to be the unit circle in . Let
be the (discrete) gradient vector at site
and be the normalized version of
(so that ). At site , we define
to be the true but unknown principal gradient direction. The (5)
normalized gradient corresponds to the observed gradient di-
rection at each site ; may not be equal to due to noise in the
fingerprint image but will be filtered with appropriate weights where
to arrive at an estimate of . The directional field at site , ,
is the direction perpendicular to , namely (6)

(1) and . We discuss the choice of

weights here. The number is the overall precision constant
Also, let , ,
which measures the contribution of the likelihood term in
and denote, respectively, the collection of
estimating . The relative precision weights correspond to
all site-wise principal gradient, directional field and observed
weighting the contributions of each block in the likelihood
normalized gradient vectors.
formula (5). We choose the weights to be
For two orientation vectors and in , the function

(2) (7)
1360 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 13, NO. 10, OCTOBER 2004

where is the coherence measure of block ; in (7),

( ) is the maximum (minimum) eigenvalue of the matrix
. We have with
indicating that the all gradient vectors in block point in
the same direction as opposed to being uniformly distributed
( ). Thus, blocks with higher (lower) values of coher-
ence will be more (less) favored in the overall likelihood, thus
providing a strong argument for this choice of . The weights
represent the contribution of each site to in block
. We choose so that the sites with large Fig. 2. Directional field for singularities based on templates: (a) core and (b)
(small) values of will have larger (smaller) contributions delta.
in (6). Note that each in (6) is normalized; if the matrix
norm of a symmetric matrix , , is its largest eigenvalue, s where . The Markovian nature of the posterior distri-
we have for all . A motivation for using bution enables simultaneous updates of disjoint coding sets (see
the density in (5) in light of 1) and 2) at the beginning of this [29], [30]) instead of updating a single site each time. This en-
section is given in Appendix I. tails significant reduction in computational time.
The block-wise maximum likelihood estimate of in (5) is We investigate the properties of the MAP estimator for some
given by the eigenvector corresponding to the maximum eigen- values of in (0,2]. In the case when , each conditional
value of . Consequently, the estimated directional field cor- update entails maximizing the function
responds to the direction orthogonal to this eigenvector. In other
words, the estimated directional field has the same direction as (13)
the eigenvector corresponding to the minimum eigenvalue of
. This is precisely the estimate of the directional field used in
with respect to . In this case, the estimate of is given by the
[2], [5].
unit eigenvector corresponding to the maximum eigenvalue of
For fixed and , the posterior density function of is given
the weighted matrix
by

(14)

(10)
We also consider the value of in this paper. When
, the MAP estimator of can be found by maximizing the
The maximum a posteriori (MAP) estimate of is obtained by function
maximizing the posterior density function in (10) with respect
to . For fixed and , this is equivalent to maximizing the (15)
objective function
with respect to . We perform this maximization numerically
(11)
in MATLAB.

We describe an algorithm to find the MAP estimator of . III. DETECTION OF SINGULARITIES

The posterior distribution of is Markovian in nature. The
conditional distribution (or, local characteristics) of given its
neighbors is
(12)
We propose to maximize the posterior distribution of in
(10) using the iterative conditional modes (ICM) algorithm [29],
[30]. Briefly, the ICM algorithm is an iterative procedure that
maximizes the conditional distribution of given its neighbors
and cycles through all sites until convergence. Each
update of the ICM algorithm increases the value of the posterior
distribution in (10), and, hence, convergence is only guaranteed
to a local maximum. Starting from an initial estimate of , , the
ICM algorithm updates each by maximizing the local char-
acteristic in (12) using the most current values of the remaining
Singularities will be detected based on comparing the ex-
tracted directional field (using the methodology presented in
Section II) with the directional field specified by parametric
template models. We use the template models for the direc-
tional field in a neighborhood around singularities as given in [2]
(see Fig. 2). The reason for using parametric template models is
two-fold: a fewer number of spurious detections are obtained
compared to nontemplate based methods, and parametric tem-
plates help mold the directional field values around detected sin-
gularities.
Let denote a neighborhood of sites centered at
. For a singular point of type
centered at and rotated degrees with respect
to the horizontal axis, the directional field vector at a site
is given by
(16)
DASS: MARKOV RANDOM FIELD MODELS 1361

where with with respect to , and ; in (22), is as defined in

and (19), is the threshold for detecting singularities, and
is a prespecified constant. The objective function in (22) is a
(17) weighed combination of information from the logarithm of the
likelihood of given , the smoothness model assumption on
and
elements of , and the pattern of the directional field around
neighborhoods of singularities based on template models with
(18) a penalty term . The parameter represents the contribu-
tion of parametric directional field forms based on template
singularity models to the overall objective function in (22);
where and are the polar representations of large values of indicate that the extracted directional field will
and , respectively. closely follow the parametric forms specified by the template
A singularity is deemed present at site if the value of the models. The penalty term allows only terms that
function exceed to be detected as singularities. One advantage of this
(19) approach is that the detected singularities mold the directional
field updates in such a way so that other singularities may
is large; in (19), is the extracted directional field vector at be detected at subsequent iterations. We illustrate this fact in
site , is as defined in (16), is the number of Section V.
sites in , and is as defined in (2). Note that, in practice, The maximization of (22) can be achieved using an iterative
the rotation angle is not known and has to be estimated. We scheme that comprises of the following conditional feature ex-
choose the estimate of that maximizes for each traction steps: 1) extract the principal gradient directions given
-template model, that is the information on singularities and 2) extract all singularity
information given the principal gradient directions. We iterate
and set (20) steps 1) and 2) and terminate when no change is observed in the
extracted features. An initial estimate of the principal gradient
(21) direction followed by a singularity extraction, can be
obtained using the methodology described in Sections II and III.
The details pertaining to the estimation of and are given in
At the th iteration ( ), we denote the extracted principal
Appendix II.
gradient directions and the singularity information by and
The value of represents the best value of similarity
, respectively. Step 1) at the -st step entails up-
of the extracted directional field with the directional field spec-
dating the values of to given . This is achieved
ified by the -template model rotated at angle with respect to
conditionally by maximizing the function
the horizontal axis. We evaluate for all sites
in a fingerprint image. The value of and are obtained
for each . The maximum of and
is then determined, and compared to a prespecified threshold
where . A singularity is said to be present at if (23)
this maximum is greater than , with singularity type and ori- using the ICM algorithm as described in Section II. Given
entation taken to be the ones corresponding to the maximum of , the singularity information is updated by maximizing
and . If the maximum is less than , we say the function
that no singularity is detected at .

IV. SIMULTANEOUS EXTRACTION OF THE

DIRECTIONAL FIELD AND SINGULARITIES
For a fingerprint image with singularities, denote by ,
, and to be the location, type (either core or delta), and
orientation of the th singularity, for . Singu-
larity information for the th singularity is denoted by the triplet
, and let . We develop
a joint feature extraction algorithm for the singularities and the
principal gradient directions, instead of the directional field, by
the equivalence in (1). Our approach will be based on maxi-
mizing the objective function
(22)
with respect to and . This can be achieved using the
methodology presented in Section III, resulting in .
Steps 1) and 2) are repeated until convergence.
V. EXPERIMENTAL RESULTS
The methodology presented in the previous sections were val-
idated on the NIST 4 fingerprint database [31]. The NIST 4
database contains 2,000 8-bit gray scale fingerprint image pairs.
Each image is 512 512 pixels with 32 rows of white space at
the bottom and classified into one of the following five classes:
arch, left loop, right loop, tented arch, and whorl. The database
is evenly distributed over each of the five classifications with
400 fingerprint pairs from each class. The results of the experi-
ments were obtained using computer code written in MATLAB.
1362 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 13, NO. 10, OCTOBER 2004

Fig. 3. Effects of  and N N

. (a)  = 0, = first order. (b)  = 1, N = first order. (c)  = 1, N = 3 2 3. (d)  = 1, N = 5 2 5. Panel (a) is Rao’s
estimate of the directional field that is commonly used.

The code was run on a PC with a Pentium IV processor with

2.00-GHz processing speed.

A. Robust Directional Field Extraction

We investigate the performance of the directional field ex-
traction algorithm presented in Section II. The value of
was fixed at 1 and the initial estimates of , , was taken
to be the MAP estimate when ; is also the max-
imum likelihood estimate of and is orthogonal to Rao’s
estimate of the directional field [5]. Fig. 3(a)–(d) illustrates the
smoothness introduced in the directional field estimates when
and are increased. Three choices of are taken: the
first order (east, west, north, and south neighbors), 3 3 and
5 5 neighborhood structures. A smoother and more robust
directional field estimate is obtained as either or the neigh-
borhood size or both are increased [compare Fig. 3(b)–(d)
to Rao’s estimate of the directional field in Fig. 3(a)]. We
also investigated the effect of different combinations of ( )
Fig. 4. Effects of increasing  and b. (a)  = 1, b = 5. (b)  = 1, b = 11.
values on the extracted directional field for fixed and . (c)  = 2, b = 5. (d)  = 2, b = 11. The image size is 296 260. The 2
We chose to be the 3 3 neighborhood structure and processing times (in seconds) are (a) 78.5, (b) 16.8, (c) 1.5, and (d) 0.5.
. In general, increasing the blocking size or or both
tend to smooth out important curvature information [compare
Fig. 4(d) with Fig. 4(a)–(c)]. Directional field extraction with
small values of are more robust with respect to increases
in [compare Fig. 4(a) and (b)] compared to larger values of
[Fig. 4(c) and (d)]. However, the reconstructions based on
takes significantly more time (see the details presented
at the bottom of Fig. 4). In all the experiments, the maximum
number of iterations was fixed at 5. Iterations above 5 did not
change the reconstructions significantly. We also compared
our results with the reconstructions reported in [2] using the
two-dimensional Gaussian kernel filter with standard devia-
tion . Fig. 5 shows the reconstruction based on various
choices of . Comparing Figs. 4 and 5, we see that better
reconstructions are obtained using the Markov random field
model. We also report the optimal blocking size (to
be discussed in Section V-D) of the fingerprint impression
presented in Fig. 4(a)–(d), namely .

Fig. 5. Directional field extraction using a Gaussian kernel filter with (a)  =
B. Singularity Detection 2, (b)  = 5, (c)  = 10, and (d)  = 20.
Fig. 6 illustrates the extraction of singularities in a fingerprint
impression based on the methodology presented in Section III. to be the optimal blocking size (to be discussed in Sec-
The directional field used to evaluate the function in tion V-D) for this fingerprint impression. The window is
(21) is obtained as in Section II, using the (3 3) neighborhood taken to be a 11 11 neighborhood around , and is taken
structure with . The blocking size was chosen to be 0.85. Fig. 6(a) shows the original fingerprint impression
DASS: MARKOV RANDOM FIELD MODELS
Fig. 6. Detection of singularities in fingerprint images based on template
models. (a) Original image. (b) Detected singularities ( represents a core and
 represents a delta). (c) Contour map of f^(C ; 1). (d) Contour map of f^(D ; 1).
with two singularities [core and delta at (row, column) (230,
300) and (370 390), respectively]. Applying succes-
sively for each point in the image results in a smooth surface
with local maximums whose locations correspond to detected
singularities of type [see Fig. 6(c) and (d)]. Thresholding re-
sults in a spatial clustering of points whose centers correspond
to locations of singularities. Separate hierarchical clustering for
is performed to determine the center of each cluster
as the location of the singular points. Fig. 6(b) shows the ex-
tracted core and delta with estimated orientation angles 28.3
and 0.48 , respectively, with respect to the horizontal axis.
C. Simultaneous Directional Field and Singularity Extraction
Obtaining directional field and singularity information simul-
taneously offers the additional advantage of dynamic updating
of features. The directional field can be dynamically molded
based on current singularity information to detect other singu-
larities in the fingerprint impression. We present two examples
here to illustrate the performance improvement of the joint fea-
ture extraction algorithm compared to extraction of the direc-
tional field followed by singularity detection. For the joint fea-
ture extraction, we took , , and the 3 3
neighborhood structure for . Fig. 7(a) gives the results when
the directional field is estimated first followed by singularity de-
tection. The singularities are detected using the filters
with fixed at 0.86. Note that one core point [at (row, column)
(180, 290)] is missed. Fig. 7(b) shows the detected singular-
ities based on simultaneous directional field and singularity ex-
traction (Section IV) using the directional field and singularity
information of Fig. 7(a) as the initial estimates ( and , with the true image gradients at singularity locations as well
respectively). Table I (second column) gives the value of [C;
(180, 290)] for three successive iterations, with the missed core
successfully detected after the first iteration.
1363
Fig. 7. Comparison of feature extraction based on two methods. (a), (c)
One-step directional field extraction followed by singularity detection. (b), (d)
Iterative updating of the directional field and singularities.
TABLE I
TABLE SHOWING THE INCREASE IN THE CORE FILTER f^(C ; 1) VALUES
The singularities in Fig. 7(c) and (d) were detected using a dif-
ferent threshold, . Note that the core at (row, column)
(250, 260) is missed when joint extraction is not performed
[Fig. 7(c)]. Fig. 7(d) shows all singularities that are successfully
detected when using the joint feature extraction algorithm. Note
the increase in [C; (250, 260)] above after the first iteration
in Table I (third column). The maximum number of iterations in
the simultaneous feature extraction was fixed at 5. No changes
were observed in the feature extractions for subsequent itera-
tions.
D. Effect of the Blocking Size,
Figs. 8 and 9 give examples of joint feature extraction based
on different block sizes . Values of that are small cause spu-
rious patterns to appear (see top rows of Figs. 8 and 9) while
large values (bottom rows of Figs. 8 and 9) level out important
ridge curvature information. Therefore, it is important to select a
block size for optimal extraction of features for each fingerprint
image. We present a criteria for selecting the optimal block size
based on similarity measures of the extracted directional field
as at pre-specified locations in the fingerprint image. Denoting
the locations of singularities by (for detected
singularities) and the location of pre-specified points on the
1364 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 13, NO. 10, OCTOBER 2004

Fig. 8. Extraction of the directional field and singularities with different block
sizes. Top, middle, and bottom rows correspond to block sizes of 5, b , and
20, respectively. The symbols and  represent detected cores and deltas,
respectively. The processing times (in seconds) are (a) 70.2, (b) 71.1, (c) 28.8,
(d) 28.2, (e) 3.4, and (f) 3.1.
Fig. 9. Extraction of the directional field and singularities with different block
sizes. Top, middle, and bottom rows correspond to block sizes of 5, b , and
20, respectively. The symbols and  represent detected cores and deltas,
respectively. The processing times (in seconds) are (a) 69.7, (b) 69.4, (c) 8.9,
(d) 49.6, (e) 3.1, and (f) 3.2.

fingerprint image by , the similarity measure with

block sizes is given by 10, 12, 14, 16, 18, and 20, and the criteria in (24) was evalu-
ated to find for each fingerprint image. A total of 25
(24) equally spaced points were chosen in the central portion of the
where fingerprint image. These constituted the pre-specified points
for . We took , for
and for , giving more
(25) weight to the fit at the pre-specified locations. Middle rows of
Figs. 8 and 9 show examples of the extracted directional field
and and singularities based on . Several spurious singularities
were detected outside of the fingerprint image. We performed a
(26) manual segmentation and recorded the number of spurious and
missed singularities within each fingerprint impression for
5, , and 20. The results are presented in Table II. The rate of
where are constants representing the penalty terms spurious detection is significantly reduced when using , al-
for spurious detection and oversmoothing, respectively, though the rate of missed singularities is increased slightly. Note
and are predetermined weights, that the rates of spurious detection based on manual segmenta-
and all other symbols are as described before. The optimal block tion presented here is significantly lower compared to those re-
size is chosen to maximize (24). ported in [2].
We applied our procedure to 1000 images from the NIST 4 Fig. 10 presents the results of simultaneous directional field
database. Nine choices of were made, namely, 5, 6, 8, and singularity extraction for five fingerprint classes [1] based
DASS: MARKOV RANDOM FIELD MODELS 1365

TABLE II models are used for singularity extraction. Joint extraction

AVERAGE NUMBER OF SPURIOUS DETECTIONS AND MISSED
SINGULARITIES PER FINGERPRINT IMPRESSIONS
of the directional field and singularities has the advantage
of dynamic updating of features and the ability to detect
previously missed singularities. One drawback is that a false
singularity point detected in the first iteration of the joint
update will be reinforced in the subsequent iterations. This is
what happens when the blocking size chosen is so small that
many false singularities are detected due to spurious ridge
patterns. The problem can be solved to some degree by choosing
the blocking size to be on an image-by-image basis.
The resulting algorithm is shown to satisfactorily extract the
directional field and singularities in a variety of fingerprints
involving moderate noise levels.

APPENDIX I
MOTIVATION FOR THE LIKELIHOOD MODELS OF
In the case of blocking with block sizes , the likelihood
term based on for block is given by

A single block contains alternating ridge and valley struc-

tures which cause the gradient vectors to fluctuate in sign. Sup-
pose that with and is either or
for . Thus, we assume that the ridge structures have one
common underlying principal gradient direction (and hence one
value of the directional field) but the sign of the gradients fluc-
tuates due to the presence of alternating ridges and valleys. We
have

with equality if, and only if, . Thus, the maximum like-
lihood estimate of , is either or and is independent
of the sign fluctuations, . Also, note that is unique up to
the sign of only. In other words, does not distinguish be-
Fig. 10. Robust extraction of the directional field and singularities for noisy tween ridge directions that are opposite to each other. It follows
fingerprint impressions for the five fingerprint classes. (a) Left loop. (b) Right that each is uniquely determined for
loop. (c) Whorl. (d) Tented arch. (e) Arch (fingerprints of class arch do not have
any cores and deltas). The processing times (in seconds) are (a) 27.7, (b) 49.2, only. The same result would hold true if the
(c) 49.9, (d) 28.3, and (e) 49.8. The values of b are (a) 8, (b) 6, (c) 6, (d) 8, function in (2) is replaced by the more general function
and (e) 6. in (9). Also, the posterior of given the observed gradients is
a function of and , and hence neighboring but opposite
on . The figures illustrate the robust extraction of the direc- directions will reinforce each other instead of cancelling each
tional field and the associated singularity detection in the pres- other out.
ence of noise. Note the robust estimation of the directional field
and detection of singularities at regions with white patches in APPENDIX II
Fig. 10(a) and (c), particularly. FINDING USING A LEAST-SQUARES CRITERIA
For two directional field vectors at site ,
, the squared directional field (SDF) vector
VI. CONCLUSION
is defined as
A class of spatially smooth models are proposed for the robust
extraction of the directional field, and parametric template (27)
1366 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 13, NO. 10, OCTOBER 2004

say, for , 2. The closeness of the two directional field ACKNOWLEDGMENT

vectors can be measured using the function in (2); we have
The author would like to thank A. K. Jain for his guidance and
his comments made during the preparation of this manuscript.

REFERENCES
(28)
[1] E. R. Henry, Classification and Uses of Fingerprints. London, U.K.:
Routledge, 1900.
from the identity and (27). In other [2] A. M. Bazen and S. H. Gerez, “Systematic methods for the computation
words, the closeness of the two directional fields can also be of the directional fields and singular points of fingerprints,” IEEE Trans.
measured in terms of the closeness of the SDF vectors. Pattern Anal. Machine Intell., vol. 24, pp. 905–919, July 2002.
[3] C. L. Wilson, G. T. Candela, and C. I. Watson, “Neural network fin-
In order to detect a singularity at , we compare the SDF gerprint classification,” J. Artif. Neural Networks, vol. 2, pp. 203–228,
values in a neighborhood of sites with that of the 1994.
template models. For , the rotated SDF template [4] L. O’Gorman and J. V. Nickerson, “An approach to fingerprint filter
design,” Pattern Recognit., vol. 22, no. 1, pp. 362–385, 1987.
at a point are given by [5] A. R. Rao, A Taxonomy for Texture Description and Identifica-
tion. New York: Springer-Verlag, 1990.
(29) [6] L. Hong, Y. Wan, and A. K. Jain, “Fingerprint image enhancement: Al-
gorithm and performance evaluation,” IEEE Trans. Pattern Anal. Ma-
and chine Intell., vol. 20, pp. 777–789, Aug. 1998.
[7] A. K. Jain, L. Hong, S. Pankanti, and R. Bolle, “An identity authetication
system using fingerprints,” IEEE Trans. Pattern Anal. Machine Intell.,
(30) vol. 85, pp. 1365–1388, Sept. 1997.
[8] N. K. Ratha, S. Chen, and A. K. Jain, “Adaptive flow orientation based
respectively. Denote the extracted directional feature extraction in fingerprint images,” Pattern Recognit., vol. 28, pp.
field and the corresponding SDF at site by 1657–1672, 1995.
[9] O. Nakamura, K. Goto, and T. Minami, “Fingerprint classification by
and directional distribution patterns,” Syst., Comput., Control., vol. 13, no.
, respectively. 5, pp. 81–89, 1982.
To determine if a core is present at , we measure the [10] V. S. Srinivasan and N. N. Murthy, “Detection of singular points in fin-
gerprint images,” Pattern Recognit., vol. 25, no. 2, pp. 139–153, 1992.
closeness of the extracted SDF with that of the template core [11] A. R. Rao and R. C. Jain, “Computerized flow field analysis: Oriented
model (29) using the least squares criteria texture fields,” IEEE Trans. Pattern Anal. Machine Intell., vol. 14, pp.
693–709, July 1992.
[12] P. Perona, “Orientation diffusions,” IEEE Trans. Image Processing, vol.
7, pp. 457–467, Mar. 1998.
[13] A. K. Jain, S. Prabhakar, L. Hong, and S. Pankanti, “Filterbank-based
fingerprint matching,” IEEE Trans. Image Processing, vol. 9, pp.
846–859, July 2000.
[14] D. Geman and S. Geman, “Stochastic relaxation, Gibbs distribution and
(31) the Bayesian restoration of images,” IEEE Trans. Pattern Anal. Machine
Intell., vol. 6, pp. 721–741, June 1984.
In (31), the rotation angle is unknown and can be chosen as [15] T. Simchony, R. Chellappa, and Z. Lichtenstein, “Relaxation algorithms
for map estimation gray-level images with multiplicative noise,” IEEE
the value that minimizes (31) with respect to . The solution Trans. Inform. Theory, vol. 36, pp. 608–613, May 1990.
is given by [16] J. L. Marroquin, S. Mitter, and T. Poggio, “Probabilistic solution of ill-
posed problems in computational vision,” J. Amer. Stat. Assoc., vol. 82,
no. 397, pp. 76–89, 1987.
[17] M. L. Comer and E. J. Delp, “Parameter estimation and segmentation of
(32) noisy or textured images using the EM algorithm and mpm estimation,”
in Proc. IEEE Int. Conf. Image Processing, vol. 2, 1994, pp. 650–654.
[18] J. Zhang, J. W. Modestino, and D. A. Langan, “Maximum likelihood pa-
rameter estimation for unsupervised stochastic model-based image seg-
It follows from (28) that mentation,” IEEE Trans. Image Processing, vol. 2, pp. 404–420, Apr.
1994.
[19] H. Derin and W. Cole, “Segmentation of textured images, using Gibbs
(33) random fields,” Comput. Vis. Graph. Image Process., vol. 32, pp. 72–98,
1986.
[20] Y. Zhang, M. Brady, and S. Smith, “Segmentation of brain mr images
and so in (32) also maximizes with through a hidden Markov random field model and the EM algorithm,”
. IEEE Trans. Med. Imag., vol. 20, pp. 45–57, Jan. 2001.
[21] D. Geman, S. Geman, C. Graffigne, and P. Dong, “Boundary detection
Using similar arguments, the value of that minimizes the by constrained optimization,” IEEE Trans. Pattern Anal. Machine Intell.,
function in (31) for a delta [from (30)] is given by vol. 12, pp. 609–628, July 1990.
[22] D. Geman and G. Reynolds, “Constrained restoration and the recovery
of discontinuities,” IEEE Trans. Pattern Anal. Machine Intell., vol. 14,
pp. 767–783, Mar. 1992.
(34) [23] P. Green, “Bayesian reconstructions from emission tomography data
using a modified EM algorithm,” IEEE Trans. Med. Imag., vol. 9, pp.
84–93, Jan. 1990.
[24] S. Geman and S. McClure, “Statistical methods for tomographic image
and we have . reconstruction,” Bull. Int. Stat. Inst., vol. 4, pp. 5–21, 1987.
DASS: MARKOV RANDOM FIELD MODELS
[25] S. Z. Li, Markov Random Field Modeling in Image Analysis. Tokyo,
Japan: Springer, 2001.
[26] R. Chellappa and A. Jain, Eds., Markov Random Fields: Theory and
Applications. New York: Academic, 1993.
[27] S. T. Acton and A. C. Bovik, “Nonlinear image estimation using piece-
wise and local image models,” IEEE Trans. Image Processing, vol. 7,
pp. 979–991, July 1998.
[28] , “Piecewise and local image models for regularized image restora-
tion using cross-validation,” IEEE Trans. Image Processing, vol. 8, pp.
652–665, May 1999.
[29] J. E. Besag, “Spatial interaction and the statistical analysis of lattice sys-
tems (with discussions),” J. R. Stat. Soc. Ser. B Stat. Methodol., vol. 36,
pp. 192–236, 1974.
[30] , “Statistical analysis of nonlattice data,” The Statistician, vol. 24,
pp. 179–195, 1975.
[31] NIST: 8-Bit gray scale images of fingerprint image groups (FIGS) [On-
line]. Available: http://www.nist.gov/srd/nistsd4.htm
1367
Sarat C. Dass received the M.Sc. and Ph.D. degrees
in statistics from Purdue University, West Lafayette,
IN, in 1995 and 1998, respectively.
He was a Visiting Assistant Professor (from 1998
to 2000) in the Department of Statistics, University
of Michigan, Ann Arbor. In 2000, he joined the De-
partment of Statistics and Probability, Michigan State
University, East Lansing. He has been actively col-
laborating with computer scientists on several pattern
recognition and image processing problems. His re-
search and teaching interests include statistical image
processing and pattern recognition, shape analysis, spatial statistics, Bayesian
computational methods, foundations of statistics, and nonparametric statistical
methods.