Professional Documents
Culture Documents
net/publication/2929466
CITATIONS READS
124 2,141
2 authors, including:
Lin Hong
Michigan State University
39 PUBLICATIONS 7,748 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Lin Hong on 24 November 2014.
Abstract
Classification
8 (k);
pixels, as the orientation of the nearest ridge seg- where
ment. Interpolate the local orientation at all pix-
els in O0 , where the corresponding pixels in A are
< if j (k )j < =2;
(k) = : + (k); if (k) ?=2;
background pixels.
? (k); otherwise;
6. Return the percentage of the area of the foreground (k) = O0 ( x(i0 ); y (i0 )) ? O0 ( x (i); y (i));
regions in A with respect to total area of A as the
quality index. i0 = (i + 1)modN :
The size of the closed digital curve is crucial for the
An example of ridge verification is depicted in Figure 7,
performance of a singular point detection algorithm us-
which demonstrates that a better orientation field can be
ing the Poincare index. If it is too small, then a small
obtained by using our ridge verification algorithm.
perturbation of orientations may result in spurious sin-
gular points being detected. On the other hand, if it is too
2.4. Singular Point Detection large, then a true pair of core and delta which are close to
one another may be ignored because the Poincare index
In an orientation field, the Poincare index of a core- of a digital curve that includes an equal number of cores
shaped singular point has a value of = and the(1 2) and deltas is 0. We have developed a singular point de-
Poincare index of a delta-shaped singular point has a tection algorithm which uses a closed square curve with
( 1 2) () ()
value of ? = [8]. Let x and y represent the x a length of 25 pixels. We have empirically determined
and y coordinates of a closed digital curve with N pix- that a curve of 25 pixels is a good trade-off between de-
( )
els. The Poincare index at pixel i; j which is enclosed tections and misses of singular points. Let O0 be the
by the digital curve can be computed as follows: interpolated orientation field. The main steps in our sin-
gular point detection algorithm are as follows:
X
N
Poincare(i; j ) = 21 (k); 1. Initialize A, which is a label image used to indicate
k=0 the singular points.
circular segments, multi-recurring segments, spiral seg-
ments, etc: The presence of a particular type of ridges
defines the class of a fingerprint. If the ridge type can be
Core
Core accurately determined, then the fingerprint can be cor-
rectly classified.
Core We classify ridges into three categories: i non- ()
Delta Delta ( )
recurring ridge, ii type-1 recurring ridge, and iii ( )
Delta
type-2 recurring ridge. Let r = fr1 ; r2 ; :::; rn g be a
ridge of length n, where r1 is the first ridge pixel, rn is
the last ridge pixel, and each pair of consecutive pixels
Figure 8. Singular point detection. is eight connected. Then r0 = fr1 ; r2 ; :::; rm g, where
m b n? c, is obtained by sampling r at intervals of
=
length . Define the cumulative orientation of r as:
( )
2. For each pixel i; j in O0 , compute the Poincare
index and assign the corresponding pixel in A a mX ?1
value 1 if the Poincare index is (1/2) and a value 2 AO(r) = j 2 1 $(k)j
if the Poincare index is (-1/2).
8 (k);
k =2
3. Find each connected component in A with pixel
< if j(k )j < ;
$(k) = : 2 + (k); if (k) ?;
values 1. If the area of the connected component
is larger than 7, a core is detected at the centroid
2 ? (k); otherwise;
of the connected component. If the area of the con- (k) = #(k) ? #(k ? 1);
where #(k ) represents the angle from r0 (k ) to r0 (k +1).
nected component is larger than 20, then two cores
Define any sequence of ridge pixels in r, fri ; r2 ; :::; rj g,
are detected at the centroid of the connected com-
where 1 i < j n, a sub-ridge of r. A non-
ponent.
4. Find each connected component in A with pixel recurring ridge, r, is a ridge such that the cumulative
values 2. If the area of the connected component orientation of any sub-ridge of r is less than a threshold,
is larger than 7, a delta is detected at the centroid Tnon = 150 o. A type-1 recurring ridge, r, is a ridge
of the connected component. such that the cumulative orientation of any sub-ridge of
r is between the two thresholds, Tnon and Trec = 270
o.
5. If more than two cores or more than two deltas are A type-2 recurring ridge, r, is a ridge such that the cu-
detected, smooth the orientation field O0 and go mulative orientation of any sub-ridge of r is larger than
back to step 1. a threshold, Trec or a ridge such that there exist mul-
tiple disjoint sub-ridges of r, which are type-1 recur-
Although the heuristic that at most two cores and two ring ridges. Obviously, uni-modal ridge segments and
deltas exist in a fingerprint is not always true, it is rarely flat ridge segments are non-recurring ridges. Circular
observed that a fingerprint has more than two cores and ridge segments, multi-recurring ridge segments and spi-
two deltas. Results of applying our singular point de- ral ridge segments are type-2 recurring ridges.
tection algorithm on two fingerprint images are shown It is very difficult to correctly extract all the true
Figure 8. ridges from an input fingerprint image, especially when
the quality of the input fingerprint image is poor. It is
2.5. Recurring Ridges essential that a ridge classification algorithm be able to
()
handle the following undesirable situations: i spuri-
The global shape of ridges determines the global con- ( ) ( )
ous ridges, ii broken ridges, and iii missing ridges.
figuration of fingerprints. Ridges in fingerprints are Ridge verification (see Figure 7) can be used to remove
highly structured. Generally, in the upper region (which all the spurious ridges from a ridge map. Broken ridges
can be roughly defined as the region above the highest can be connected based on the information present near
core points in loops, tented arches, and whorls and the the end of broken ridges. However, it is very difficult to
region above the most curved ridges in arches) of a fin- recover missing ridges. This needs both high-level struc-
gerprint, ridges are a family of uni-modal smooth curve tural analysis and local structural analysis of the ridge
segments. In the bottom region, ridges form a family of pattern, which is very difficult to formulate and imple-
relatively flat curves. In the middle region, depending ment. We have developed a ridge classification algo-
on the fingerprint class, ridges may be of the follow- rithm which traces each ridge in the verified ridge map
ing types: uni-modal curve segment, recurring segment, and classifies each ridge into one of the three categories
(a) (b) (c) arch tented arch left loop
Figure 9. Ridge classification; ridges clas-
sified as (a) non-recurring, (b) type-1 re-
curring, and (c) type-2 recurring.
fingerprints in NIST-4 database are labeled as belong- Reject rate 0% 5% 10% 20%
ing to two different classes, each row of the confusion 5-class Error 12.5% 11.6% 10.1% 7.5%
matrix in Table 1 does not sum up to 800. For the five- 4-class Error 7.7% 6.6% 5.1% 2.4%
class problem, most of the classification errors are due to
misclassifying a tented arch as an arch. By combining Table 2. Error-reject tradeoff.
these two arch categories into a single class, the error
rate drops to 7.7%. Besides the tented arch-arch errors,
the other errors mainly come from misclassifications be- Acknowledgments
tween arch/tented arch and loops and due to poor image
quality. Two examples of misclassified fingerprints are This research was supported by an IBM university
shown in Figure 11. A lower error rate can be achieved partnership program.
by adding the reject option, which is based on the quality
index of the input image. The error rates corresponding References
to different reject rates are listed in Table 2.
[1] P. Baldi and Y. Chauvin. Neural networks for fingerprint
5. Summary and Conclusions recognition. Neural Computation, 5(3):402–418, 1993.
[2] G. T. Candela, P. J. Grother, C. I. Watson, R. A. Wilkin-
son, and C. L. Wilson. PCASYS: A Pattern-Level Classi-
Fingerprint classification provides an important in- fication Automation System for Fingerprints. NIST Tech.
dexing mechanism for automatic fingerprint identifica- Report NISTIR 5647, August 1995.
tion. At a first glance, the fingerprint classification prob- [3] M. Chong, T. Ngee, L. Jun, and R. Gay. Geometric frame-
lem appears to be rather simple. But because of large in- work for fingerprint image classification. Pattern Recog-
traclass and small interclass variations in global pattern nition, 30(9):1475–1488, 1997.
configuration and poor quality of input images, the de- [4] A. Jain, L. Hong, and R. Bolle. On-line fingerprint veri-
sired accuracy of 1% error rate at 20% reject rate is very fication. IEEE Trans. Pattern Anal. and Machine Intell.,
19(4):302–314, 1997.
difficult to achieve. We have designed a fingerprint clas-
[5] K. Karu and A. K. Jain. Fingerprint classification. Pattern
sification algorithm which classifies input fingerprints Recognition, 29(3):389–404, 1996.
into five categories according to the number of singu- [6] M. Kawagoe and A. Tojo. Fingerprint pattern classifica-
lar points detected, their relative position, and presence tion. Pattern Recognition, 17(3):295–303, 1984.
of type-1 and type-2 recurring ridges. Our algorithm in- [7] H. C. Lee and R. E. Gaensslen. Advances in Fingerprint
vests a significant amount of effort in feature extraction Technology. Elsevier, New York, 1991.
to make robust to interclass variations as well as poor [8] D. Monro and B. Sherlock. A model for interpreting fin-
quality of input images. Experiment results demonstrate gerprint topology. Pattern Recognition, 26(7):1047–1055,
that our algorithm has better classification performance 1993.
[9] E. Newham. The Biometric Report. SJB Services, New
than previously reported in the literature on the same
York, 1995.
database. Currently, we are studying an alternative clas-
sification scheme which classifies fingerprints according
to the pattern similarity. We are also investigating an im-
age enhancement algorithm based on a bank of Gabor
filters.