Pattern Recognition, Vol. 23, No. 8, pp. 893 904, 1990 0031-3203/90 $3.00 + .

00
Printed in Great Britain Pergamon Press pie
© 1990 Pattern Recognition Society

AUTOMATED FINGERPRINT RECOGNITION USING

STRUCTURAL MATCHING
ANDREW K. HRECHAK and JAMES A. MCHUGH
Department of Computer and Information Science,New Jersey Institute of Technology,Newark NJ 07102, U.S.A.

(Received 7 June 1989; in revised form 1 November 1989; received for publication 20 November 1989)

Abstract--Automated fingerprint recognition has received considerable attention over the past decade.
Progress has been made on models of the structure of fingerprints, techniques for the acquisition of prints,
and the development of commercial automated fingerprint recognition systems. Despite these advances,
there remain considerable opportunities for improvement. The speed of retrieval, and the ability to
recognize partial or distorted prints are prominent among those areas that require improvement. This
study will describe a structural model of fingerprints, based on local structural relations among features,
and an associated automated recognition system which addresses the limitations of existing fingerprint
models.

Fingerprint recognition Image processing Feature extraction Structural matching

Partitioning Identification

Figure 1 shows a basic print enlarged to illustrate

l. I N T R O D U C T I O N
some common features. "5~ The definition of Galton
features and their extensions play a fundamental role
The use of fingerprints for identificationtl-7) is based in the design of automated systems. 19,16 25)
on the immutability and the individuality of finger- The most prevalent current model for automated
prints. Immutability refers to the permanent and fingerprint identification systems is the Minutiae-
unchanging character of the pattern on each finger, Coordinate model (also called the FBI representation
from before birth until decomposition after death. of prints). Most commercial systems are based on the
Individuality refers to the uniqueness of ridge details basic FBI model and an extension due to Wegstein/2~
across individuals. No two persons, even identical In this model, two types of print features, namely
twins, have been found to have identical fingerprints, breaks and forks in the ridge lines of the print,
despite elements of similarity.~g'91 are taken as the distinctive features of prints. The
Fingerprints have two levels of structure: the Henry coordinates of each feature as well as the angle with
(1900) fingerprint pattern (with ridge count) ~5~and the respect to the horizontal of a tangent to the feature
Galton (1892) characteristics/4) The Henry Classi- are recorded and the list of coordinates is used to
fication is the standard qualitative scheme for charac- represent the print.
terizing the global structure of ridge patterns and A more refined feature-based data model for finger-
has traditionally been used to partition fingerprint prints is described in a theoretical study performed
databases. Recently, automated fingerprint classi- by Ratkovic at the Rand Graduate Institute. 122) This
fication techniques have been investigated.~1°'1t~ model identified 10 different types of print features,
However, the individuality or uniqueness of a finger- based on the Osterburg study, t26~ A subsequent work
print is not based on its global Henry structure, but by Liu et al. manually recognized features using a
on the existence and position of so-called ridge grid that was overlaid on a print, and then summarized
characteristics (Galton characteristics or features) on the arrangement of features in a code. t91
the print. 14~In order to establish unique identification, Recently, the Sparrows at the National Bureau
a sufficient number of identical characteristics identi- of Standards introduced another model/23,24~ The
cally positioned must be located. Sparrow model of fingerprints uses a structural rep-
Galton originally defined four characteristics: resentation of fingerprints designed to allow the
(1) beginnings and ends of ridges, (2) forks, (3) islands, recognition of a print even in the presence of distor-
and (4) enclosures/4~ Over time, the list of characterist- tions and (small) rotations and translations. The
ics has been refined or extended/~2'~3~ The American model overlays a (horizontal or vertical) reference line
National Standards Institute has proposed an identi- on the print (through its center) and uses the features
fication standard with four types of characteristics arising on ridge lines passing through the reference
or features: ridge endings, bifurcations, compounds line to characterize the print. The representation is
(trifurcations or crossovers), and undetermined, c~4~ the basis for a related search algorithm.
893
894 ANDREW K. HRECHAK and JAMES A. MCHUGH

i. SHORT RIDGE
2.&
3. BIFURCATION
4.&
5. BIFURCATION
6. RIDGE

7. ENDING RIDGE

8.&
9. BIFURCATION

i0. RIDGE

Ii. ENDING RIDGE

12. RIDGE

13. SNORT RIDGE

14.&
15. BIFURCATION

16.&
17. ISLAND

18.&
19. BIFURCATION

20. ENDING RIDGE

21. DOT

22. RIDGE

23.&
LOOP 24. ISLAND

25 81DO~ COUNTS 25. ENDING RIDGE

Fig. 1. Sample enlarged print and ffatu~s. (is)

Current fingerprint recognition systems suffer from and type of features in the neighborhood are identified
several disadvantages. Their retrieval abilities are and counted, determining a feature vector. The model
weak if the prints are of poor quality, or are partial, for the print consists of a list of such feature vectors.
stretched or twisted, or contain falsely reported or A naive matching algorithm would be to identify a pair
missing features. The systems are sensitive to errors of prints if their collection of feature neighborhoods
in the positioning of prints at acquisition, and hence coincided. We show instead that a small number of
basically require complete prints in order to position distinctive neighborhoods can be found which can be
the prints correctly for subsequent coordinate calcu- used to uniquely characterize each print in a database.
lation or reference line placement. The structure of Furthermore, the characterizations remain relatively
the current data models for fingerprints leads to effective even when only part of a print is available.
computationally costly print comparisons at match- This structural data model directly addresses cer-
ing time (that is, when a test print is compared with tain critical drawbacks of existing systems:
the successive prints in the database in search of a (1) The model is positionally invariant. This means
closest match). The models generally entail brute- that the feature vectors obtained are the same inde-
force sequential searches and are incompatible with pendently of how the print is positioned. This property
a successive partitioning approach to the search space. indirectly improves processing speeds because the
This study will introduce techniques that are applic- matching algorithm does not have to account for
able to the recognition of clear partial prints even positionally caused errors.
when only part of the print is available, and that are (2) The model is robust in the presence of incomplete
robust with respect to translation, rotation, and data because even a fragmentary print may still
distortions. The methodology is based on a data contain the critical set of defining neighborhoods.
model for fingerprints that is structural, rather than Furthermore, even if it contains only a subset of its
coordinate. critical neighborhoods, the list of possible matching
The model uses local configurations of features to prints may be able to be significantly narrowed.
recognize and retrieve prints, and appears to support (3) The model supports more efficient search strateg-
more rapid and robust search and retrieval strategies ies since the feature vectors are relatively short; require
than current models. The model uses eight classical only trivial computations to compare; and can even
types of features. A small neighborhood is specified be searched on subvectors to allow fast elimination
around each feature on a print. Then, the number of poor candidates.
 Automated fingerprint recognition 895

Though our data model differs considerably from

current systems, the existing methodology of finger-
print recognition supports the feasibility of this
approach. For example, the model (and search pro-
cedures) emulates in certain respects what a fingerprint
expert does who manually searches for a matching
fingerprint by implicitly using local positional
relations between features to recognize a print. Our
approach can be viewed as an attempt to both
formalize and automate this technique.

2. F I N G E R P R I N T CONVERSION

The fingerprint matching algorithms were tested

experimentally using a database of fingerprints whose
development constituted the first phase of our work,
digital fingerprint conversion. This section overviews
the process used in converting an image of a print to
a compact digital representation. Figures 2-7 show
a sample fingerprint processed through the entire Fig. 3. 2. enlarged print.
conversion process. The conversion process is over- . . . . I ° ,~ o • L
viewed in Fig. 8. The steps involved include the
following:
(1)Obtain a physical print.
(2) Use a video camera with the appropriate soft-
ware to capture the print image.
(3) Store the complete print graphics image in a file
as a color image, with varying degrees of intensity
(each image was internally represented as an array of
pixeis, where each pixel was represented by three
colors (red, green, and blue) of varying intensity).
(4) Convert the digitized color image to a suitable
storage matrix representation, using thresholding to
map the color bit vectors to a 0-1 (black and white)
image.
(5) Store a compacted representation of the storage
matrix in the file (original 100K byte image com-
pressed to 6 K byte representation).
(6) Thin and clean compacted representation. This '--. ~ " ~ ' ~ t ~
entails: Fig. 4. 3, original digital print.
(a) using the classical Pavlidis thinning algor-
ithm t27~ to reduce the image ridges to a skeletal
structure,
(b) smoothing the resulting fingerprint distor-
tions, and
(c) selecting a valid fingerprint region, by deleting
regions with no identification significance, such
as those generated by overinking. The deletion
criteria were based on visual inspection.

o," y.&... •

Fig. 2. 1. physical print Fig. 5. 4. thinned print.

896 ANDREWK. HRECHAKand JAMESA. MCHuGH

Fig. 6. 5. cleaned print. Fig. 7. 6. select print region ready for feature extraction.

The features needed to test different matching for identification of fingerprints. Eight features were
algorithms were then extracted from the resulting selected, including the four types established by Gal-
image. Regions of the image which were very noisy ton, ta) and four others generally accepted by finger-
were deleted, while the remaining noisy areas were print experts and which appear on fingerprints with
handled by allowing a limited number of false patterns. non-negligible frequencies.
Recall that, while we have paid some attention to the The features were broadly classified into two
occurrence of noisy areas on the prints, our underlying classes: primitives, and compound features, which are
supposition is that the prints in the database are to combinations of primitives. The features used, in
be interpreted as clear prints, such as might be the order of their extraction by the recognition
obtained using a better acquisition system than we algorithms, are:
have used, such as laser-based input techniques, or Primitives
by definitively isolating clear areas of noisy prints. (I) Dot
(2) Ridge Ending
(3) Bifurcation
3. FEATURE EXTRACTION Using the primitives as a base, it was possible to
define the compound features.
The next phase involved the extraction of finger- Compound features
print features from the digital images produced by (4) Island
the conversion phase. There were many possible (5) Spur
choices for a set of basic individual features in a (6) Crossover
fingerprint, so we first established a standard inven- (7) Bridge
tory of features and methods for their automated (8) Short Ridge
recognition. A subset of the approximately 31 features Islands are defined as two bifurcations with two
currently recognized in the literature was selec- connecting paths. Spurs are formed by a combination
ted. {4'12'22'26) The selection was designed to ensure of a bifurcation and a ridge ending. Crossovers
both that the features had fundamental support in are defined as two or more neighboring connected
the fingerprint literature and appeared to be useful bifurcations. Bridges are defined as two bifurcations

PHYSICAL
PRINT
200>I 256

CAMERA
FORM
199

PRINT
....
199

STORAGE
FORM
STORAGE
...... >
COMPACT
PHYSICAL
MAP> MAPPING FORM

102,418 B 39,601 B 5,572 BYTES

FILE MEMORY FILE

RGB IMAGE B/W IMAGE

Fig. 8. Basic conversion process.
 Automated fingerprint recognition 897

with one connecting path. Short ridges are defined as FEATURE AVERAGE FREQUENCY
two neighboring connected ridge endings. The fea- % OF TOTAL FEATURES # PER PRINT
tures were defined in a manner which permitted well- DOT 4.3 18
defined identification for automated recognition. See
RIDGE ENDING 60.6 258
Fig. 9 for an illustration and Fig. 10 for the frequency
distribution of the features in the sample population. BIFURCATION 17.9 76

The hierarchical nature of the design allows it to be ISLAND 0.7 3

readily extended to recognize more complex types of SPUR 4.7 20
compound features. CROSSOVER 3.2 14
The features were recognized by scanning the BRIDGE 2.5 i0
fingerprint via windows. Windows of size 3 x 3 pixels,
SHORT RIDGE 6.1 26
with an overlap of two, were used to extract the
primitive features. Then combinations of these primi- Fig. lO. Feature ~equency distribution.
tives were used to derive the compound features.
The feature recognition algorithms were iteratively
For example, higher priorities were assigned to
adjusted and re-designed until all the features were
features that had greater print identifying power, so
properly recognized under varying conditions. The
features less useful for discrimination did not mask
process involved the interactive display of fingerprints
more useful features. Size and structural limits were
with given features highlighted (with colored pixels),
considered. Figure 11 illustrates the effect of size
allowing computer-aided visual verification of the
parameters on the recognition of features: 1 denotes
accuracy with which feature types and locations were
an island recognized using a window size of 9, while
recognized. The kind of issues which were identified
2 denotes an island found using a window size of 19.
during this stage included:
Since fingerprints can be rotated, translated, and
(1)the priority of extraction of features,
distorted, the robustness of our feature definition had
(2) positional invariance problems,
to be substantiated. This positional invariance of
(3) border conditions,
feature extraction was tested by selecting a random
(4) the interaction between features, and
set of prints, which were rotated through 90, 180, and
(5)choice of feature parameters (size, shape, and
270 °. (Other angles of rotation were harder to simulate
defining conventions).
effectively because of the limited resolution of our
digitized representations.) The resulting rotated fea-
VISUAL tures were identical. Translation testing also yielded
FEATURE ILLUSTRATION correct results, the features being invariant, as
expected. The correct classification of distorted fea-
tures is implicit in the topological (shape-oriented)
DOT nature of our classification scheme. That was verified
by extensive visual checking of how features were

/
RIDGE ENDING

BIFURCATION

/IT, ." I
ISLAND " ; ' < 'L",,
SPUR

CROSSOVER

BRIDGE

SHORT RIDGE
Fig. 11. Sample island window determination:
Fig. 9. Sample ~ature variations. 1--standard window; 2--expanded window.
898 ANDREW K. HRECHAK and JAMES A. McHuoH

classified and iteratively adapting the recognition feature neighborhoods (only feature frequencies arc
algorithms until the recognition was verified to be recorded, not feature positions) makes the scheme,
robust, as we have indicated previously. The accurate like the morphological approach of the Henry System,
recognition of features was again tested via the color relatively robust against geometric distortions, and
feature indicator scheme described earlier. applicable even if only partial prints arc available.
The structural data model emulates in a sense the
conceptual model of a fingerprint expert. An expert
4. FINGERPRINT MATCHING recognizes prints by looking for tell-talccorrespon-
dences of features and may well bc able to match a
The structural model for fingerprints is defined as database print against a test print even if only a
follows. F o r each feature on the fingerprint, we define fragment of the test print is available. The expert can
a neighborhood of some specified radius R about the use global properties of the print to eliminate potential
center of the feature. If we denote the allowed types matching candidates or to position and orient a
of features by T~... T8 each occurring with frequency fragment of a print against a more complete database
f(TO...f(Ts) in the neighborhood, and the type of print. Hc can then correlate features by using their
the central feature for the neighborhood by T, then type, position, orientation, and location relative to
we associate with the central feature a vector other features. Our data model implicitly captures
(T,f(TO...fiTs)) as a characterization of the neigh- much of this expert behavior, since our feature vectors
borhood. Figure 12 gives a hypothetical example of define an associative mechanism which effectively
a feature and its neighborhood, where the central simultaneously locates and correlates features.
feature is of type 1 and the feature vector is (1, 0, 1, The specific approach used was to identify a small
0, 1, 1, 0, 0, 0). With a suitable choice of R, we can set of feature vectors V~ ...Vk, whose occurrence on
obtain a list of such feature vectors for each print, a print uniquely identifiedthe print, in the context of
containing information similar to that useful in a our 150 print database. Initially,wc identified a set
manual search; namely, distinctive local patterns of of 9 feature vectors which uniquely identifiedany one
feature distribution, which can serve as the basis for of a test set of 60 prints, from a given set of 60
search and identification. fingerprints in our database. W e then extended this
This model represents a synthesis of the global set of feature vectors to 15 vectors which together
partitioning approach, typified by the Henry class- uniquely identified any one of the 150 prints in our
ification, and the local identification approach, typ- database.
ified by feature based models such as the FBI system. Our basic procedure was as follows: determine the
We will experimentally determine a set of vector feature vectors for each print; condense the list of
patterns which effectively define a partition of much feature vectors to eliminate duplicates; then perform
finer granularity than the limited number of partition a data analysis to determine vectors that appear to
types provided under a Henry scheme. The feature bc useful for identification.It was eventually decided
based nature of the classifcation makes it more that a set of feature vectors with a mix of frequencies
amenable to automation than the Henry system since of occurrence across all prints ranging from 15-75%
the features are both local in character and relatively provided a useful basis for identification.The results
sharply defined in nature. The relational nature of the of the vector selection process for five 30 print test
sets,individually and combined, are shown in Fig. 13.
The table illustrates that as the number of prints
Central Feature increases, the required number of identifying vectors
to support unique identificationappears to also slowly
increase, as suggested by the following discussion.
For example, using II vectors, three 30 print test
sets have uniquely identified prints, with only one
duplicate ID in set three and two duplicate ID's in
set four. Combining the sets, wc have eight duplicate
ID's and three ID's which identify three prints each.
Overall, the vectors used appeared on average in
approximately 4 2 % of the prints. A group of vectors
(in this case 15) which allows unique print identifi-
cation is shown in Fig. 14. Some distribution statistics
for these vectors arc shown in Fig. 15. The table
illustrates the number of prints with a given vector,
including a breakdown of those prints with the vector
occurring one, two, or three times per print. O n
average, 8 1 % of the vectors fallinto these breakdown
categories, with the less frequent vectors almost com-
Fig. 12. Hypothetical neighborhood. pletely accounted for.
 Automated fingerprint recognition 899

# OF ID VECTORS structural matching algorithm against this database.

SET 9 ii 12 14 15 The experiment considers both full and partial prints.
In the full print case, each test print was processed
1 1 1 1 1 against the database using the structural model bit
1 1 1 1 1 map technique resulting in the retrieval of a matching
2(4) 2(1) 2(1) 1 1 print. In the partial print case, each of a variety of
test partial prints was processed against the database
3(i)
using the structural model bit map matching approach
4 2(~) 2(2) 2(2) 1 1
and a ranked list of candidates was retrieved.
5. 2(1) 1 i i 1 The database used contains 150 fingerprints, whose
......... i ......................................... input and feature extraction were both automated
150 PRINT 2(22) 2(8) 2(8) 2(2) 1 as described earlier. The database population was
DATABASE 3(6) 3(3) 3(1)
designed to be a representative sample, including
prints of both low and high quality. See Fig. 16 for a
6(1)
sample database fingerprint with selected features
Fig. 13~ ID vector list determination results: # 1 ( 41: 2) marked. The entire database is represented by a
:1t: I - Max # of given ID's (1-All UNIQUE ID's); compact Fingerprint Bit Map Table along with an
4# 2 - Number of times duplicate set occurs. associated Vector List. The resulting storage require-
Each set contains 30 prints. ments are only 4 bytes per print.
We also investigated vector lists with other distribu- A test set of 50 fingerprints was selected representing
tional characteristics. We considered for example lists a random sample of prints of varying quality from
consisting of only relatively infrequent vectors (such the database. The set of complete (clear) prints was
as, each of about 10% frequency), as well as lists just the test set itself. The set of partial (clear) prints
consisting of relatively average vectors (such as, each was derived by deleting regions from the test set
of about 50% frequency). Neither of these types of prints.
distributions appeared to be as useful in identifying The results of the experiments are presented in Figs
prints as was the mixed distribution. The specific list 17-19. Depending on the type of print tested, the
of features chosen does not seem to be critical, as other percentage of test prints with a correct match ranked
sets of feature vectors with reasonable performance in the top M positions is shown. A match was
characteristics appeared to be determinable. determined by the existence of corresponding feature
vectors on a print, as indicated by the active bits in
the print bit map. Only those database prints with
5. E X P E R I M E N T A L RESULTS
the same active bits, ignoring the state of other bits,
were extracted. The matching was then ranked by the
This section provides a description of the test lowest number of extra active bits, thereby placing
database, and the results of experiments testing our print matches with a greater degree of similarity in a
higher position. Table l shows matching results for
ID VECTOR L I S T
complete prints, ¼prints, and ½prints. Table 2 shows
results for four experiments with ¼prints. Table 3
FEATURETYPES:DGCISUNE
shows the results for a ¼print which is distributed
CENTER TYPE # D FREQUENCY i 0 0 0 0 0 0 1
over two ~ print size square regions.
E 0 0 O O 1 O O 1 A summary of the results follows. By design (that
S 0 0 0 0 1 0 0 0 is, by the choice of the feature vectors), we obtain
C 0 0 0 1 0 0 0 0 100% matching for complete prints. Due to the
S 0 0 0 0 0 0 1 0 simple, computationally non-intensive nature of the
U 0 0 0 0 0 0 1 0
computations, the matching time is negligible (!). This
initial experiment demonstrates it is possible to select
D 0 0 0 0 1 0 0 1
sets of feature vectors which are intuitively meaningful,
I 0 0 0 0 1 0 0 0
while still allowing a small number of feature vector
S I 0 0 0 0 0 0 0 patterns to retrieve or uniquely identify every print
U 0 0 0 0 0 1 0 0 in the database.
S 1 0 0 0 1 0 0 1 The results using partial prints are also promising.
C 0 0 0 0 0 0 1 0 We use the standard active bit matching procedure.
D 1 0 0 0 1 0 0 0
Observe that, in the first table, a match is identified
in the top 5% of the database (that is, occurring
C 0 0 0 0 0 1 1 0
before the seventh or eighth position of the output
H 0 0 0 0 0 0 1 0
list) for over 90% of the test prints for ¼ partials and
Fig. 14. Database ID vector list: D: Dot, S: Spur; G: Ridge 80% of the test prints using ½ partials. For over 90%
Ending; U: Crossover; C: Bi~rcation; H: Bridge; I: Island; of the ½ partials the match is in the top 9% of the
E: Short Ridge. database prints. The ¼ partial results in the second
900 ANDREW K. HRECHAK and JAMES A. MCHUGH

# TIMES/PRINT TOTAL
VECTOR # # PRINTS 1 2 3 ~ OF # PRINTS

1 80 24 19 17 75
2 54 33 14 4 94

3 96 18 23 18 59
4 75 30 16 i0 56

5 84 44 27 6 92

6 63 36 18 6 95

7 46 41 3 i 98

8 28 18 9 0 96

9 76 42 17 II 92

10 58 17 23 6 79

11 21 17 4 0 i00

12 115 25 15 18 50

13 31 20 I0 0 97

14 89 22 13 ii 52

15 53 23 20 3 87

Fig. 15. ID vector list distribution statistics: ~ prints - =M:of prints from 150 database with vector;
¢e times - =M=of prints with given vector count.

SELECT PRINT REGION

..,... ~,~,<

3 . _..H

8 1
~ •

SAMPLE PRINT FEATURE SUMMARY

1 - DOTS = 18

2 - RIDGE ENDINGS = 292

3 - BIFURCATIONS 53

4 - ISLANDS = 2

5 - SPURS 17
6 - CROSSOVERS 12
7 - BRIDGES 3
8 - SHORT RIDGES = 30

Fig. 16. Sample print.

 Automated fingerprint recognition 901

TYPE OF PRINT COMPLETE 3/4 PARTIAL 1/2 PARTIAL

MATCH RANK % OF TEST PRINTS

M i I00 66 34

M2 76 56

M 3 60

M4 82 66

M5 84

M6 86 68

M 7 90 76

M8 80

M9 84

MI0

MII 86

MI2 90

MI3 92

MI4 94

MI5

Fig. 17. Matching experiment results - Table 1.

1/4 PARTIAL
PRINT RANDOM TOP MIDDLE BOTTOM

MATCH RANK % OF TEST PRINTS

MI 20 12 i0 14

M2 28 16 12 18

M3 32 24 14 24

M4 42 32 16 32

M5 46 36 20 34

M6 . . . . 24 "

M7 50 44 28 36

M8

M9 58 50 36 42

MIO " 52 38 46

MII " 56 " 48

MI2 60 58 44 "

MI3 . . . . 48 "

MI4 " 62 50 "

,, ,, ,, ,,
MI5

M~O 64 70 62 50

M39 82 80 78 74

M73 92 90 88 84

Fig. 18. Matching experiment results - Table 2.

902 ANDREWK. HRECHAKand JAMESA. McHUGH

TYPE OF PRINT RANDOM DISTRIBUTED 1/4 PARTIAL boundary effects. By allowing one false vector (that
MATCH RANK OF TEST PRINTS is, using an active search bit/non-active database bit),
the missed prints were ranked in the top 3 4 0 (that
M1 12 is, the top 51 positions) of the database. At most 2%
M2 16 (currently, one print) of all the test prints were not
M3 22 recognized using one false bit.
M4 24

M5 26
6. C O N C L U S I O N
M6 30

M7 32 The objective of this study was to develop a

MS data model for fingerprints which: allowed correct
M9 38 recognition of prints, was robust in the presence of
MIO 46 positioning errors, was applicable to partial prints,
Mll 50
and supported fast retrieval and matching. We have
defined a structural model of fingerprints, based
MI2
on local structural relations among features, and
MI3
developed an associated automated recognition sys-
MI4 tem* which addresses the limitations of existing
MI5 fingerprint models.
There are a number of directions in which the
M2o 52
research could be extended.
Mi9 66
(1) The most obvious extension is to a larger-scale
database. It would be useful to develop the database
M~3 78 using a laser print reader system with a suitable
digitized output.
Fig. 19. Matching experiment results - Table 3. (2) The model has intentionally avoided any reliance
on purely geometric (coordinate) information. How-
ever, if positional information about a print or print
fragment is known, it would be appropriate to allow
table vary slightly depending on whether fixed or a general system to utilize this information.
random regions are used, with the random regions (3) Another extension is to allow the special neigh-
performing slightly better. A match is ranked in first borhoods to be characterized not only by the number
place for an average of 20% for the random partials and types of their features, but also by the numbers
(vs 12% for the fixed ¼ partials). A match ranking of ridge lines in the neighborhood.
within 5% of the database occurs for 50% of the (4) An enhanced set of features could be developed.
random partials (vs 36% for the fixed ¼ prints). For (5) The matching technique could be made more
over 60% of the test prints, it is possible to get a robust against noise by adapting the system to recog-
matching candidate within 9% of the database even nize neighborhoods missing some of their defining
with a randomly selected ¼print. Finally, for all the features, such as might occur in a noisy environment.
¼ trials, a match can be ranked within 50% of the (6) Maximum use should be made of the information
database prints for about 90% of the test prints. In available in a latent print. In our current system, the
the third table, we consider random ¼ partials which latent is processed by the automated feature extraction
have been split into two random segments each equal subsystem. An alternative is to design a computer-
to ~ of a print with results similar to those for the aided interface that facilitates an interactive character-
fixed partials in the earlier experiment. A match is ization of the features of a latent.
ranked in first place for 12% of the test prints, within
5% of the database prints for 32% of the test prints,
and within 7% of the database prints for 5 0 0 of the SUMMARY
test prints. At this point a match can be ranked within
50% of the database prints for about 80% of the test This study described a new structurally based
prints. automated fingerprint recognition system that uses
Since, it may happen that a partial print, either local patterns of features to identify prints.
because of noise or boundary effects, may generate a The software development of the system was initi-
false pattern which is not a true feature of the print, ated by the design of a digital conversion process for
we guard against such occurrences by extending the fingerprints, whose ultimate output was a fingerprint
matching technique to allow a limited number of false ridge structure ready for feature extraction. Then,
matching vectors. For example, 8% of the test ¼
partials were missed (that is, there was no proper
match in the active bit matching list) because of * Patent Pending.
 Automated fingerprint recognition 903

extraction algorithms were developed which permit- 8. C. H. Lin, J. H. Liu, J. W. Osterberg and J. D. Nicol,
ted the automated recognition of eight classic finger- Fingerprint comparison. I: Similarity of fingerprints, J.
print features useful for discriminating among prints. Forensic Sci. 27, 290-304 (1982).
9. J. H. Liu, C. H. Lin, J. W. Osterberg and J. D. Nicol,
Feature tables were developed for each print, which Fingerprint Comparison. II: On the development of a
made it possible to examine various alternative struc- single fingerprint filing and searching system, J.
tural models based on local relations among features. Forensic Sci. 27, 305-317 (1982).
A digitized database of 150 representative finger- 10. B. Moayer and K. S. Fu, A syntactic approach to
fingerprint pattern recognition, Pattern Recognition 7,
prints was used to test the model under a variety of
1-23 (1975).
scenarios based on the nature of the search prints. 11. K. Rao and K. Balck, Type classification of fingerprints:
Despite the limited nature of this database, this testing a syntactic approach, IEEE Trans. Pattern Anal. Mach.
environment was comparable to or superior to the lntell. PAMI-2, 223-231 (1980).
test environments used for most other published 12. K. D. Saviers, Friction Skin Characteristics: A Study
and Comparison of Proposed Standards, Garden Grove
research efforts in automated fingerprint recognition. California Police Department, July (1987).
The experimental results substantiated the usefulness 13. D. A. Stoney and J. I. Thornton, A critical analysis of
of this technique for automating fingerprint recogni- quantitative fingerprint individuality models, J. Forensic
tion. For example, the experiments included testing Sci. 31, 1187-1216 (1986).
the ability to retrieve a print when two disjoint parts 14. Fingerprint Identification-Data Format for Information
Interchange. American National Standards Institute,
of the print, together covering only 25% of the area New York (1986).
of the complete print, were available. 15. The Science of Fingerprints: Classification and Uses.
The system described herein addresses some notable Federal Bureau of Investigation, U.S. Government Print-
difficulties in the area of automated fingerprint recog- ing Office, Washington, D.C. (1984).
16. Application briefs: Computer graphics in the detective
nition, including the limited ability to recognize business, I E EE Comput. Graphics Appl., pp. 14-17, (April
partial or distorted prints, and the considerable com- 1985).
putational cost required to attain a reasonable rate 17. D. K. Isenor and S.G. Zaky, Fingerprint identification
of retrieval even when good prints are available, using graph matching, Pattern Recognition 19, 113-122
(1986).
caused by the computationally intensive nature of
18. B. Reed, Automated fingerprint identification:from Will
existing fingerprint models. The structural model West to Minnesota Nine-Fingers and beyond, J. Police
of fingerprints we have defined and its associated Sci. Admin. 9, 317-326 (1981).
automated recognition system address these limi- 19. The state of development of the FBI's automatic finger-
tations. Significantly, our model allows prints to be print identification system: Part I, FBI Law Enforcement
Bull., pp. 8-13, (June 1973).
compared even when only partial prints are available 20. The state of development of the FBI's automatic finger-
because the model representation of a partial print is print identification system: Part II, FBI Law Enforcement
a subset of the model representation of the whole Bull., pp. 8-13, (June 1973).
print. The comparison of a pair of prints on this 21. J. H. Wegstein, An Automated Fingerprint Identification
system involves a negligible a m o u n t of time, since it System. National Bureau of Standards Special Publi-
cation 500-89, U.S. Government Printing Office, Wash-
entails only a simple bit vector comparison. The cost ington, D.C. (1982).
of processing is basically subsumed in the (one-time 22. J. A. Ratkovic, Increasing Efficiency in the Criminal
only) cost of preprocessing each print to extract its Justice System: The Use of New Technology for Criminal
select neighborhood structure, while the comparisons Identification and Latent Print Processing. The Rand
Corporation, California, (September 1980).
entailed during run-time searches are relatively trivial. 23. M. K. Sparrow and P. J. Sparrow, A Topological
Notable elements of the system include automated Approach to the Matching of Single Fingerprints: Devel-
recognition of c o m p o u n d features, and the identifi- opment of Algorithms for Use on Latent Fingermarks.
cation of local feature patterns with significant National Bureau of Standards Special Publication 500-
retrieval power. 126, U.S. Government Printing Office, Washington, D.C.
(1985).
24. M. K. Sparrow and P. J. Sparrow, A Topological
REFERENCES Approach to the Matching of Single Fingerprints: Devel-
opment of Algorithms for Use on Rolled Impressions.
1. C. E. Chapel, Fingerprinting---A Manual of Identifi- National Bureau of Standards Special Publication 500-
cation. Coward McCann, New York (1941). 124, U.S. Government Printing Office, Washington, D.C.
2. J. F. Cowger, Friction Ridge Skin: Comparison and (1985).
Identification of Fingerprints. Elsevier, New York (1983). 25, T. F. Wilson and P. L. Woodard, Automated Fingerprint
3. Fingerprint Identification. Federal Bureau of Investi- Identification Systems: Technology and Policy Issues.
gation, U.S. Government Printing Office, Washington, Bureau of Justice Statistics, U.S. Government Printing
D.C. (1977). Office, Washington, D.C. (1987).
4. F. Galton, Finger Prints. Macmillan, London (1892). 26. J. W. Osterberg, T. Parthasarathy, T. E. S. Raghaven
5. E. R. Henry, Classification and Uses of Finger Prints. and S. L. Sclove. Development of a mathematical formula
Routledge, London (1900). for the calculation of fingerprint probabilities based on
6. W. J. Herschel, Skin furrows of the hand, Nature 23, 76 individual characteristics, d. Amer. Statis. Assoc. 72, 772-
(1880). 778 (1977).
7. w. R. Scott, Fingerprint Mechanics A Handbook. 27. T. Pavlidis, Algorithms for Graphics and Image Process-
Charles C. Thomas, Springfield, Illinois (1951). ing. Computer Science Press, Rockville, MD (1982).
904 ANDREW K. HRECHAK and JAMES A. MCHUGH

About the Author--ANDREW K. HRECHAK received the B.S. and M.S. degrees in Computer Science from
New Jersey Institute of Technology in 1980 and 1982, respectively. He obtained the M.B.A. and Ph.D.
degrees from Rutgers University in 1988. His research interests lie in the application of computer
technology in criminology. Currently he is associated with the New Jersey Institute of Technology.
Dr Hrechak is a member of ACM, IEEE, Alpha Iota Delta, Beta Gamma Sigma, Phi Eta Sigma, Tau
Beta Pi, and Upsilon Pi Epsilon.

About the Author--JAMES A. MCHUGH received the B.S. degree in Mathematics from Fordham College
in 1965, and a Ph.D. in Mathematics from the Courant Institute of Mathematical Sciences in 1970.
At present, he is a Professor in the Department of Computer and Information Science at the New
Jersey Institute of Technology. His current research interests are algorithmic graph theory, analysis of
algorithms, and image processing.