Oewey
HD28
.M414
MIT LIBRARIES DUPL
39080 00932 7575
Algorithms
for
Thinning
and
Rethickening Binary
Digital
Pattern
M.V.
NagendraprasadPatrick
Wang
Amar
Gupta
WP#3764
1993Productivity
From
Information
Technology
(PROFIT)
The
Productivity
From
Information
Technology
(PROFIT)
Initiativewas
established
on
October
23,
1992
by
MIT
President
Charles
Vest
and
Provost
Mark
Wrighton
"tostudy
the
use
of
information
technology
in
both
the private
and
public
sectors
and
to
enhance
productivity
in
areas
ranging
from
finance
to
transportation,
and
from
manufacturing
to
telecommunications."
At
the
time
of
itsinception,
PROFIT
took
over
the
Composite
Information
Systems
Laboratory
and
Handwritten
Character
Recognition Laboratory.
These two
laboratories are
now
involved
in
research
re-lated
to
context
mediation
and
imaging
respectively.
./.as&achuseits instituteOF TECHNOLOGY
MAY
2
3 1995
LIBRARIES
In
addition,
PROFIT
has
undertaken
joint efforts
with
a
number
of
research
centers,
laboratories,
and programs
atMIT,
and
the
results
of
these
efforts
are
documented
in
Discussion Papers published
by PROFIT
and/or
the
collaborating
MIT
entity.
Correspondence can be addressed
to:The
"PROFIT"
InitiativeRoom
E5
3-3
10,
MIT
50 Memorial
Drive
Cambridge,
MA
02142-1247
Tel:(617)
253-8584
Fax:
(617)
258-7579
E-Mail:
profit@mit.edu
EXECUTIVE
OVERVIEW
Financial enterprises
relyheavily
on
paper-based
documents
toconduct
various
operations;
this istrue
both
forexternal
operations
involving
customers
and
other
financial institutions,as
well
asinternal
operations involving various
departments.
Researchers
atMIT
have
looked
atthe
possibilityof
taking information
directly
from
paper documents,
especially
handwritten documents,
tocomputer-accessible
media.
Automated
reading involves
several steps as follows:
(i)
Scanning
of
document;
(ii)Location
of
area
tobe
"read";(iii)
Decomposing
the selected
area into
separate
characters;
(iv)
Adjusting
sizeand
slantof
each
character;
(v)
Recognizing each
character;
and
(vi)
Testing
whether
input
has
been
correctly read.
Based
on
several
years
of
sustained
research, the researchers
have
attained
very
high
"reading"
speed
and
accuracy,
even
insituations
where
the quality of the
input
material
ispoor.
Patent
rights forsome
of the
new
techniques
have
been
applied
for.Sponsor
companies
are
eligible to testthe
new
techniques
in theirrespective
environments
atno
charge.
The
work
performed
so
far isdescribed
ina
number
of
published
paper
and
working
papers.
The
listof
working
papers
isas follows:
IFSRC
#
107-89 OpticalImage
Scanners
and
Character
Amar
Gupta
Recognition
Devices:A
Survey
and
New
Sanjay
Hazarika
Taxonomy
Maher
KallelPankaj Srivastava
IFSRC
#
123-90R
An
Improved
StructuralTechnique
for PatrickS. P.Wang
Automated
Recognition
ofHandprinted
Symbols
Amar
Gupta
Revised
October
1990
IFSRC
#
124-90 IntegrationofTraditionalImaging, Expert
Evelyn
Roman
Systems,
and
Neural
Network
Techniques
forAmar
Gupta
Enhanced
Recognition
ofHandwritten
John
Riordan
Information
IFSRC
#
151-91Handwritten
Numeral
Recognition
Using
Ronjon
Nag
Dynamic Programming
Neural
Networks
on an
AlexisLui
Off-Line
BasisAmar
Gupta
IFSRC
#
162-91R Algorithms
forThinning
and
Rethickening
M.
Nagendraprasad
PROFIT
93-03
Binary
DigitalPatterns Patricks.Wang
Amar
Gupta
Handwritten
Characters
IFSRC
#
214-92An
Algorithm
forSegmenting Handwritten
Numeral
StringsIFSRC
#
21 5-92A New
Algorithm
forCorrectingSlant in
Handwritten Numerals
TFSRC
#
218-92A
System
forAutomatic
Recognition
ofTotally
Unconstrained
Handwritten Numerals
IFSRC
# 219-92
A
CoUection
ofPapers
on Handwritten
Numeral
Recognition
IFSRC
# 261-93
An
Adaptive
Modular
Neural
Network
with
Application
toUnconstramed
Character
RecognitionIFSRC
# 287-94
An
Integrated ArchitectureforRecognition
of
PROFIT 93%4
TotaUy
Unconstrained
Handwritten Numerals
IFSRC
# 288-94
Detection
ofCourtesy
Amount
Block
on
Bank
PROFIT
93-09
Checks
IFSRC
# 289-94
A
Knowledge
Based
Segmentation Algorithm For
PROFIT
94
14
Enhanced R^ognition
ofHandwritten
Courtesy
Amounts
Amar
Gupta
Peter L.Sparks
M.
V.Nagendraprasad
Amar
Gupta
M.
V.Nagendraprasad
Amar
Gupta
Vanessa
FelibertiM.
V.Nagendraprasad
Amar
Gupta
LikMui
Arun
Agarwal
P. S. P.Wang
Amar
Gupta
M.
V.Nagendraprasad
A.Liu
Amar
Gupta
S.Ayyadurai
Arun
Agarwal
Len M.
Granowetter
Amar
Gupta
P. S. P.Wang
Karim
Hussein
Amar
Gupta
Arun
Agarwal
Patrick
Shen-Pei
Wang
the
papers,
and
the
software
developed
atMIT.
The
Principal Investigator
forthe
imaging
area
isDr.
An>f
Gupta
Co-Directo'^-PROF,^-
IniHaavl
MIT
^'-^^^3^
^-^^-^-^.^^^'^^^^^^^
^^^^;
fSf
-a^uTSt-r
S'c^e^^d
suggestions
are
DIGITALSIGNAL PROCESSING3,97-10211993)
Algorithms
for
Thinning
and
Rethickening
Binary
Digital
Patterns
M.
V.
Nagendraprasad,
Patrick
S. P.Wang,* and
Amar
Gupta^
MassachusettsInstitute ofTechnology, Cambridge, Massachusetts
02139
1. INT
RODUCTION
Pattern recognition
and
image processingapplica-tions frequently deal with
raw
inputs that containlinesofdifferentthickness. In
some
cases, thisvaria-tioninthe thicknessis
an
asset,enablingquickerrec-ognition of the featuresinthe input image. For
exam-ple, in processing aerial photographs, detection of
major
landmarks
canbeaidedbythe variationsinthethickness of the contours. In other cases, the
varia-tion can be a liability,
and
can cause degradation intheaccuracy
and
the speedof recognition. Forexam-ple, in the case ofhandwrittencharacters, thedegree
of uniformity ofthe thickness of individual strokes
directlyimpactsthe probability of successful
recogni-tion, especially if neural
network
based recognitiontechniques areemployed.
For the latter category of applications, uniform thickness can beattained, priorto recognition stage,
by firstthinningthe input pattern to a thickness of a
single pixel
and
then rethickening it to a constant thickness.The
basic structureand
the connectivity of the original pattern can be preserved irrespective ofthe underlying complexity, through the stages of
thinning
and
rethickening.Digitized
bitmap
patterns consist ofan
array ofpixels,
where
each pixel is either 1 ("on" pixel) or("ofT' pixel). In thinning,alsocalledskeletonization, the
redundant "on"
pixels are eliminatedfrom
theoriginal pattern to yield its equivalent skeletonized pattern.
During
thesubsequentstage of rethickening,•CollegeofComputerScience, NortheasternUniversity.
'Thisresearchwasfunded by
the International Financial
Ser-vices Research Center at MIT's Sloan School of Management.
Comments andsuggestions should be addressedtothe principal
investigator for this project: Dr. Amar Gupta,
Room
E53-311,MIT,Cambridge,
MA
02139,USA;telephone(617) 253-8906."on"
pixels are systematicallyadded
to reconstructan
equivalent of the original pattern. Because thethinning process isusually considered
more
difficultthan the rethickening process, thebulkofthispaper dealswith thinningaspect.
Section 2 deals with basic notation.
The
thinningstage is discussed in Section 3. Section 4 presents a theoreticalproofrelatedtoa
new
and
fasterthinning algorithm.The
rethickening stageisdiscussedinSec-tion5. Results arepresentedinSection 6
and
conclu-sions in Section 7.2.
BAS
ICNOTATION
One
ofthe authors [11] has previously presenteddefinitions
and
notation relatedtothethinningalgo-rithms presented here. In order to facilitate a direct
comparison
of thenew
algorithm with a previousoneproposedin [11],the
same
notation isutilized in thispaper.
Definition1.
The
neighborsof apixel,p:[i,)],areidentifiedbythe eightdirections, [i
-
l,j], [i—
l,j+
1], [',;"
+
1], [i+ hj
+
1], [i+
IJ],[i+
IJ
-
1], [ij -1], [i-
I,j-
1).The
directions are also assigned anumber
k taking valuesfrom
0, . . ., 7 asshown
inFig. 1.
Definition2.
The
contourpoints of adigitalpat-tern aredefined as thosepixels for
which
at leastone neighbor is off. InFig. 2,"a,""b," . . . , "k"and
some
ofthepixels 'm'and
'n'arecontourpoints whilenone
ofthe"1"s isacontourpoint.
Definition3.
The
contourloopisasetofcontourpoints
which
are connected into a loop.More
for-mally,asetofcontourpointsc,,Cj,. . .,c„(for(n
>
1)
form
a loopiffc,isaneighborofc,+,forK
i<
n and
c„isaneighborofc,.
We
use L(l),. . . ,L{m)
tolabelthe1051-2004/93 $4.00 Copyright
©
1993byAcademicPress,Inc.The
parallel thinning algorithm byWang
and
Zhang
in [11] performed thinningasfollows:Algorithm
WZ:
initial;
g=
1;repeat
<?
=
<?1;if
(g==
1)theng=
0:elseg=
1;for
;:=
1to
flidoif
73[;f] = 1then
begin
p
=FIRST[;f]
;x
= PREV[;c] ;h[k]
=
0;repeat
z
=
successor
{x, p) ;x
=
p;case
goi
0:
ifl<B(p)
<7
and (A(p)
= 1or
c(p)
=
1)and
(p[2] +
P[4])*p[0]*
P[6]
=0
then
begin
deletion
{Ql. p)h[k]
=
1end;
1:if
1<B(p)
<7and
(i4(p) = 1or
c(P)
=
1)and(
p(0] +
P[6])
*Pl2]* p[4]
=0
then
begin
deletion
(^1, p);
h[k]
=
1end;
end
case;
P=
z;until
loop-end-test
{k)end
until h[l] +
• •+ h[m]
= 0.The
proposednew
parallelalgorithmoperates asfol-lows:
Algorithm
NWG:
g=
1;repeat
h=
0:Q =
Q1:if
g==
1then
g=
0;else
g=
1;for each
array element p
of
^
begin
if contour(
p)then
case
gof
0:if
1< S(p) <
7
and
(>l(p) = 1or
c{p)
=
1)and
(p[2]
-I-P[4])*p[0]*
p[6]
=0
then
begin deletion
(^1, p) ;h[k]
=
1; A=
1;end;
1:if
1< 5(p) <
7
and
(>l(p)=
ior
c{p)
=
1)and
{p[0] +
P[Q])*p[2]*
Pl4]
=0
then
begin deletion
(^1 , p) ;
h[k]
=
1;A =
1;end;
end
case;
end
for;until
h=
0;In each iteration, the
new
algorithm uses only a subset of the operations involved in theearlieralgo-rithm.
During
asinglepass over the full array,Algo-rithm
NWG
usestwo operations—
contour(p)and
de-letion (Ql, p), whereas Algorithm
WZ
used initial,loop-end-test(k), successor(x, p),
and
deletion (Q, p),FIRST[*],
PREV(/e)and
contour(p) as a part ofini-tial, FIRST[/e]
and
PREV[)ij]. Further, the functioncontour(p)isusedat leastas
many
timesinFIRST[/e]asitisinAlgorithm
NWG;
at leastone
scanisneededtodetect the starting points ofallthecontour
loopsin
the array.
The
speed ofthenew
algorithm issignifi-cantly higherthantheearlierone. Alsothe
new
algo-rithm can be
programmed
more
readily.JUNCTIONAL
EQUIVALENCE
PROOF
This section provides a proofthat the
new
algo-rithm is operationally equivalent to the
earlier
algo-rithm, that identical intermediate outputs occur at
the
end
ofeachiteration for bothcases,and
that thefinaloutputs will alsobe identical.
Lemma
1. Ifthedata arrayatthebeginningofithpass'isidentical forboth the algorithms,
and
further ifAlgorithm
WZ
marks
a pointp
for deletionduring thepass in the data array, then Algorithm
NWG
alsomarks
thesame
pointp
for deletion.Proof For Algorithm
WZ,
tomark
pointp
forde-letion, the following conditions
must
besatisfied:• point
p
must
beon one
of them
contour loops;and
• thecasestatement
must
have been satisfied.But
ifthe pointp
isa part ofany
contourloop, thencontour(p)
must
besatisfied. Ifcontour(p)issatisfiedand
the case statement is satisfied, then AlgorithmNWG
willmark
the pointp
for deletion.Lemma
2. //the data arrayatthebeginningofithpassisidentical forboth the algorithms,
and
further ifAlgorithm
NWG
marks
a pointp
for deletion duringthepass in the data array, then Algorithm
WZ
alsomarks
thesame
pointp
for deletion.Proof
For
AlgorithmNWG,
tomark
pointp
fordeletion, the following conditions
must
besatisfied:'Hereapassmeansthenumber
of times
Q
= Ql sUtement isexecuted. Ifthisalgorithm ismodifiedforexecutionona parallel
computer, thenthe step involving
Q
=Qlwillnotexist.Thenthenumberofpasses over the arrayisthenumberoftimes
Q
=QlgetsIfa point
p
ismarked
outfordeletionby
AlgorithmNWG,
then the following conditionsmust
besatis-fied:
• contour (p)
must
be true;and
• the case statement
must
have been satisfied.But
ifcontourip) istrue,thenp
must
belongtoone
ofthe
m
contoursand
hencemust
beone
ofthe pointsencountered while Algorithm
WZ
traverses thecon-tourloops.
When
thispointisencounteredand
ifthecase statement is satisfied, then Algorithm
WZ
willmark
the pointp
fordeletion.Lemma
3. AlgorithmWZ
and
AlgorithmNWG
ex-ecute
an
identicalnumber
ofpasses over the dataarray.Proof.
The
proofis obviousfromtheway
thetwo
algorithms are constructed.
During
apass. AlgorithmWZ
goes overallthecontoursinthe arrayonceand
ifthere isa deletion duringthe traversal of
any
of theiiiiiii iiiiiiii iiiiitiii iiitiii 1111 mil till iiiiii mil mill tim mil
mi
mil 1111 mil mil mil minimmiiiim
miiiiiii 111 mil mil mil mill iiiiiii miiiii iimiii iiiiiim miiiii 11111111 mini mini mill mil imiiii iimiiit iimiiiiimi
111 111 nil nil nilmi
iinmniimi
mini nil miiiii 111 imniiiii iimiiiim
mini nil mil 111 miiiiii imniiii miiiiin 111 nilnmi
mini millmm
m
111 nil mil mill ill 111 111 mil nnnniiii iniiiii n 111 nil mil milmi
mi
nil nil nil nil 11111 nit minniinni nmiiiiiininmmiini
111 nil 111 nil nil ncontour points, then there is another iteration over the data array. Algorithm
NWG
goes over ail the pointsinthe arrayand
deletesonly points thatresideon
the contour. If there isany
deletion during thepass,thenit goesthrough another iteration.Sinceat
the
end
ofany
intermediate iteration, thesame
out-put isproduced by
bothofthem, they willterminateafter
an
identicalnumber
ofiterations.Theorem.
AlgorithmNWA
and
AlgorithmWZ
are operationally equivalent, that is, given the
same
input, theyproduce the
same
output.Proof. Follows
from
Lemmas
1, 2,and
3 above.5.
RETHICKENING
The
skeletonization algorithm abovewas
devel-oped
as a part of a larger ongoing projecton
auto-mated
handwritten numerals.The
numerals
inputtoa generalhandwriting recognition system could have
beenwritten using variouskindsofpenslikefeltpens,
ball pens, ormicrotippedpens.
Each
ofthese writinginstrumentscreatescharacters ofdifferentthickness.
In order toreducethese variations,theinput pattern
is scaled
down
to a single pixel thicknessand
thenrethickened to a uniform thickness.
The
algorithmabovethins
down
thenumeral
patterntoa thicknessof
one
pixel. This"skeletal"bitmap
ofthenumeral
isnow
rethickened to a standard thickness, allowingnumerals
ofvaryingthickness to be rethickenedtoauniformthickness.
Even
withinthebitmap
array of anumeral,the thickness of the
numeral
may
bediffer-entin different parts.
Thinning
and
rethickeningcanalso
smoothen
these variations to a large extent.The
algorithmforrethickening looksatapre-spec-ified neighborhood of each l_pixel
and
fills thisneighborhood with I's.
6.
RESULTS
The
National InstituteofStandardsand
Technol-ogy
(NIST) Handprinted
CharacterDatabase
was
utilized as the basis for providing standardized
in-puts.
The
raw
datainabitmap
form,withan
averagesizeofabout 50
X
350,were normalizedtoa 16X
150array.Inthelatterarray,thedata
were
typically4 to6pixels in thickness. Figure 3
shows
asample
set ofnumerals
before skeletonizationand
thecorrespond-ingskeletonized
numerals
obtained as the output oftheproposedalgorithm.Figure4
shows
afewsamplesofdigits
which were
skeletonizedand
then rethick-ened.Note
that the digits inthe first column,which
7.
CONCLUSION
MIT LIBRARIES
39080 00932 7575
Apart frompresenting themotivationfor thinning
and
rethickening,thispaper hasdiscussedalgorithmsfor performingthese tasks efficiently.
The new
thin-ning algorithm has been
shown
tobefaster,butfunc-tionally identicalto,one of the fastestthinning
algo-rithms in existence.
By
using the algorithmspre-sented in this paper, raw inputs with underlying
deficiencies intermsofunequalthicknesscan be
nor-malized to yieldhigheroverall speed
and
accuracy.REFER
ENCES
1. Adairid.M. R..Martin, G..andSuen,C. Y., Tracing center-linesofdigitalpatterns usingSquare and
Maximal-CircleAlgorithms. Proc. of VisionInterface' 90. Halifax, Can-ada, 1990, pp.67-79.
2. Arcelli,C,Cordelia, L.P.,andLevialdi,S..Fromlocalmaxima toconnectedskeletons.
IEEE
Trans. Pattern Anal. Mach. In-(e«. 3(1981), 134-143.3. Arumugam,A., Radhakrishnan, T.,andSuen,C.Y.,
A
thin-ning algorithmbased ontheforcechargedparticleandits
ex-perimental comparison. Personalcommunication.
4. Davies,E.R.,Plummer,A. P. N.,Thinningalgorithms:
A
cri-tique and a new methodology. Pattern Recognition 14. 1(1981),53-63.
5. Lu, H.E.,andWang,P. S.P.,
A
commenton"Afast parallelalgorithmfor thinningdigital patterns." Commun.
ACM
29(1986),239-242.
6. Naccache,N.J.,andShinghal,R.,Aninvestigationinto
skele-tonization approach of Hilditch. Pattern Recognition 17. 3 (1984),274-284.
7. Pavalidis, T.,
A
flexible parallelthinning algorithm. InPro-ceedings of IEEE ComputerSociety Conference on Pattern Recognitionand Image Processing, 1981, pp. 162-167.
8. Sossa,J.H.,Animprovedparallelalgorithmforthinning digi-talpatterns.Pattern RecognitionLett.(1989),77-80. 9 Stefanelli,R.,andRosenfeld,A.,Someparallelthinning
algo-rithmsfor digital pictures.J.
ACM
18(1971),255-264.10. Tamura, H.,
A
comparisonofline thinning algorithmsfromdigitalgeometryviewpoint.Inthe Proceedingsof4th
Interna-tionalConference on Pattern Recognition, 1978, pp.715-719.
11. Wang,P. S.P.,and Zhang.Y.Y.,
A
fastandflexiblethinning algorithm.IEEE
Trans. Comput.38.5 (1989).M.V.
NAGENDRAPRASAD
obuinedhisB.Tech fromIndianInstitute ofTechnology, Madras, and M.S. degreesfrom Indian
InstituteofTechnology, Madras, and Massachusetts Institute of
Technology.Heiscurrentlyastudentatthe Universityof
Massa-chusetts,Amherst, working towardsadoctorateinDistributed
Ar-tificial Intelligence. His interesU include artificial intelligence,
connectionist systems,machinelearning,andcognitivescience.
PATRICK
S. P.WANG
isaProfessorofComputerScienceatNortheasternUniversity. He received his Ph.D. degree in
com-puter sciencefrom OregonState University,hisM.S.I.C.S. degree
from Georgia InstitueofTechnology, his M.S.E.E. degree from
National Taiwan (.'niversity,and his B.S.E.E. degree from
Na-tionalChiaoTungUniversity.
Before joining Northeastern University, he had been with
WANG
Laboratories,GTE
Laboratories, BostonUniversity,andthe UniversityofOregonassoftware engineeringspecialist,senior
memberoftechnicalsUflF,adjunct AssociateProfessor,and
Assis-tantProfessor, respectively.From 1989to 1990,hewasavisiting scientistatthe
MIT
Artificial IntelligenceLaboratory.Heisnowalsoteaching part-timeatHarvard University,andisa research consultantatthe
MIT
Sloan SchoolofManagement.Professor
Wang
has published sixbooks andover 70technicalpapers in imaging technology, pattern recognition,and artificial intelligence.Oneof hismostrecent inventions concerning pattern recognitionsystemshasbeengrantedapatentbythe U.S. Patent Bureau. He has attendedandorganized numerousinternational conferencesforIEEE,
ACM,
ICS,AAAI, CLCS,andtheInterna-tional AssociationforPattern Recognition. Heisalso areviewer
forgrants proposalsofthe
NSF
IntelligentSystemsDivision,andfor several computerjournals including
lEEEPAMI.
ComputerVision, Graphics,and ImageProcessing,InformationSciences,and
Information ProcessingLetters,and aneditor-in-chargeofthe In-ternationalJournal of Pattern Recognition and Artificial Intelli-gence.
Prof.
Wang
isaseniormemberofIEEE, andalso amemberofACM,
AAAI. andthe Pattern Recognition Society. Dr.Wang
ischairman oflAPR-SSPR(1992-1994) andhasbeena governing
boardmemberandthe newslettereditoroftheChinese Language
ComputerSociety(CLCS), Washington.
DC
(1983-1989). Hisre-searchinterests include software engineering,programming lan-guages, pattern recognition,artificialintelligence,image
technol-ogy,andautomation.
Professor
Wang
hasalsowrittenseveral articlesonthe operasof Verdi, Puccini, and Wagner, and on Beethoven's, Mozart's andTsaikovsky'ssymphonies.
AMAR
GUPTA
(SM'85) is Senior Research Scientist at theSloanSchool ofManagement.MassachusettsInstitute of
Technol-ogy,Cambridge.Heholdsabachelor's degreein electrical
engineer-ingfromthe Indian InstituteofTechnology, Kanpur, a master's degreein managementfromtheMassachusetUInstitute of
Tech-nology, anda doctorate incomputertechnologyfrom the Indian
Institute ofTechnology,
New
Delhi.Heconductedthe researchfor hisdoctoratedissertation atthreeprestigious universities in India,England, andtheUnitedStates.
Dr.Guptahasbeeninvolvedinresearch,management,and
pub-lishing activities since 1974. Since joining
MIT
in 1979, he hasbeen active in the areas ofmultiprocessor architectures,
perfor-mancemeasurement,distributedhomogeneousandheterogeneous
data bases, expert systems, personal computers,and transfer of information technology. Heserves as aconsultantto anumberof corporations, government agencies, and international bodies on various aspects ofcomputertechnology,andisanactivetechnical advisorto severalinternational organizationsandcommittees.
Dr GuptaistheChairmanofthe TechnicalCommitteefor
Mi-croprocessor ApplicationsoftheIEEEIndustrialElectronics Soci-ety and assistant chairman for several annual
lECON
confer-ences.Hewasalsoinvolved with theVery Large Data Base Confer-ence(VLDB
'87),heldinEnglandin 1987.Hereceived theRotary Fellowship for International Understanding in 1979 and theBrooksPrize(Honorable Mention)in 1980).
Hehas writtenmorethan 50technicalarticlesandpapers,and producedseven books.
Dr.Guptaisa citizenof theUnitedStates.