ulglLal lmage processlng ls Lhe use of compuLer algorlLhms Lo perform lmage

processlng on dlglLal lmages As a subcaLegory or fleld of dlglLal slgnal processlng

dlglLal lmage processlng has many advanLages over analog lmage processlng lL
allows a much wlder range of algorlLhms Lo be applled Lo Lhe lnpuL daLa and can
avold problems such as Lhe bulldup of nolse and slgnal dlsLorLlon durlng
processlng Slnce lmages are deflned over Lwo dlmenslons (perhaps more) dlglLal
lmage processlng may be modeled ln Lhe form of MulLldlmenslonal SysLems
Many of Lhe Lechnlques of dlglLal lmage processlng or dlglLal plcLure processlng
as lL ofLen was called were developed ln Lhe 1960s aL Lhe !eL ropulslon
LaboraLory MassachuseLLs lnsLlLuLe of 1echnology 8ell LaboraLorles unlverslLy
of Maryland and a few oLher research faclllLles wlLh appllcaLlon Lo saLelllLe
lmagery wlrephoLo sLandards converslon medlcal lmaglng vldeophone
characLer recognlLlon and phoLograph enhancemenL1 1he cosL of processlng
was falrly hlgh however wlLh Lhe compuLlng equlpmenL of LhaL era 1haL
changed ln Lhe 1970s when dlglLal lmage processlng prollferaLed as cheaper
compuLers and dedlcaLed hardware became avallable lmages Lhen could be
processed ln real Llme for some dedlcaLed problems such as Lelevlslon sLandards
converslon As generalpurpose compuLers became fasLer Lhey sLarLed Lo Lake
over Lhe role of dedlcaLed hardware for all buL Lhe mosL speclallzed and
compuLerlnLenslve operaLlons

WlLh Lhe fasL compuLers and slgnal processors avallable ln Lhe 2000s dlglLal
lmage processlng has become Lhe mosL common form of lmage processlng and
generally ls used because lL ls noL only Lhe mosL versaLlle meLhod buL also Lhe
cheapesL

ulglLal lmage processlng Lechnology for medlcal appllcaLlons was lnducLed lnLo
Lhe Space loundaLlon Space 1echnology Pall of lame ln 19942

1asks

ulglLal lmage processlng allows Lhe use of much more complex algorlLhms for
lmage processlng and hence can offer boLh more sophlsLlcaLed performance aL
slmple Lasks and Lhe lmplemenLaLlon of meLhods whlch would be lmposslble by
analog means

ln parLlcular dlglLal lmage processlng ls Lhe only pracLlcal Lechnology for
ClasslflcaLlon
leaLure exLracLlon
aLLern recognlLlon
ro[ecLlon
MulLlscale slgnal analysls

Some Lechnlques whlch are used ln dlglLal lmage processlng lnclude
lxellzaLlon
Llnear fllLerlng
rlnclpal componenLs analysls
lndependenL componenL analysls
Pldden Markov models
AnlsoLroplc dlffuslon
arLlal dlfferenLlal equaLlons
Selforganlzlng maps
neural neLworks
WaveleLs
lxellzaLlon ls a vldeo and lmageedlLlng Lechnlque ln whlch an lmage ls blurred
by dlsplaylng parL or all of lL aL a markedly lower resoluLlon lL ls prlmarlly used for
censorshlp 1he effecL ls a sLandard graphlcs fllLer avallable ln all buL Lhe mosL
baslc blLmap graphlcs edlLors
A famlllar example of plxellzaLlon can be found ln Lelevlslon news and
documenLary producLlons ln whlch vehlcle llcense plaLes and faces of suspecLs aL
crlme scenes are rouLlnely obscured Lo malnLaln Lhe presumpLlon of lnnocence
as ln Lhe Lelevlslon serles CCS 8ysLanders and oLhers who do noL slgn release
forms are also cusLomarlly plxelaLed looLage of nudlLy (lncludlng genlLalla
buLLocks or breasLs) ls llkewlse obscured ln some medla before Lhe waLershed ln
many counLrles ln newspapers or general magazlnes or ln places ln whlch Lhe
publlc cannoL avold seelng Lhe lmage (such as on blllboards) urug references as
well as gesLures consldered obscene (such as Lhe flnger) may also be censored ln
Lhls manner lxellzaLlon ls noL usually used for Lhls purpose ln fllms uvus
subscrlpLlon Lelevlslon servlces pornography (excepL for counLrles ln whlch Lhe
law requlres lL) When obscene language ls censored by an audlble bleep Lhe
mouLh of Lhe speaker may be plxelaLed Lo prevenL llp readlng

lxellzaLlon may also be used Lo avold unlnLenLlonal producL placemenL or Lo
hlde elemenLs LhaL would daLe a broadcasL such as daLe and Llme sLamps on
home vldeo submlsslons Censorshlp for such purposes ls mosL common on reallLy
Lelevlslon serlesclLaLlon needed

edlL lxellzaLlon ln medla

lxellzaLlon has also been used for arLlsLlc effecL noLably ln Lhe arL prlnL 1he
Wave of Lhe luLure a relnLerpreLaLlon of kaLsushlka Pokusals 1he CreaL Wave aL
kanagawa ln Lhls updaLed prlnL Lhe lmage of Lhe large ocean wave shlfLs from
Lhe LradlLlonal sLyle of Lhe !apanese woodcuL prlnL Lo a plxelaLed lmage and
flnally Lo a wlreframe model compuLer graphlcs lmage12 WesLworld (1973)
was Lhe flrsL feaLure fllm Lo use dlglLal lmage processlng Lo plxelaLe phoLography
Lo slmulaLe an androlds polnL of vlew plxellzaLlon ls used ln Lhe slmswhen Lhe
charlcLor does someLhlng Lo do wlLh nudlLy (llke baLhlng or uslng Lhe LolleL)

Llnear fllLers ln Lhe Llme domaln process Llmevarylng lnpuL slgnals Lo produce
ouLpuL slgnals sub[ecL Lo Lhe consLralnL of llnearlLy 1hls resulLs from sysLems
composed solely of componenLs (or dlglLal algorlLhms) classlfled as havlng a llnear
response


MosL fllLers lmplemenLed ln analog elecLronlcs ln dlglLal slgnal processlng or ln
mechanlcal sysLems are classlfled as causal Llme lnvarlanL and llnear Powever
Lhe general concepL of llnear fllLerlng ls broader also used ln sLaLlsLlcs daLa
analysls and mechanlcal englneerlng among oLher flelds and Lechnologles 1hls
lncludes noncausal fllLers and fllLers ln more Lhan one dlmenslon such as would
be used ln lmage processlng Lhose fllLers are sub[ecL Lo dlfferenL consLralnLs
leadlng Lo dlfferenL deslgn meLhods whlch are dlscussed elsewhere

A llnear LlmelnvarlanL (L1l) fllLer can be unlquely speclfled by lLs lmpulse
response h and Lhe ouLpuL of any fllLer ls maLhemaLlcally expressed as Lhe
convoluLlon of Lhe lnpuL wlLh LhaL lmpulse response 1he frequency response
glven by Lhe fllLers Lransfer funcLlon P() ls an alLernaLlve characLerlzaLlon of Lhe
fllLer 1he frequency response may be Lallored Lo for lnsLance ellmlnaLe
unwanLed frequency componenLs from an lnpuL slgnal or Lo llmlL an ampllfler Lo
slgnals wlLhln a parLlcular band of frequencles 1here are a number of parLlcularly
deslrable or useful fllLer Lransfer funcLlons of whlch Lhls arLlcle wlll presenL an
overvlew

Among Lhe Llmedomaln fllLers we here conslder Lhere are Lwo general classes of
fllLer Lransfer funcLlons LhaL can approxlmaLe a deslred frequency response very
dlfferenL maLhemaLlcal LreaLmenLs apply Lo Lhe deslgn of fllLers Lermed lnflnlLe
lmpulse response (ll8) fllLers characLerlsLlc of mechanlcal and analog elecLronlcs
sysLems and flnlLe lmpulse response (ll8) fllLers whlch can be lmplemenLed by
dlscreLe Llme sysLems such as compuLers (Lhen Lermed dlglLal slgnal processlng)

rlnclpal componenL analysls (CA) ls a maLhemaLlcal procedure LhaL uses an
orLhogonal LransformaLlon Lo converL a seL of observaLlons of posslbly correlaLed
varlables lnLo a seL of values of uncorrelaLed varlables called prlnclpal
componenLs 1he number of prlnclpal componenLs ls less Lhan or equal Lo Lhe
number of orlglnal varlables 1hls LransformaLlon ls deflned ln such a way LhaL Lhe
flrsL prlnclpal componenL has as hlgh a varlance as posslble (LhaL ls accounLs for
as much of Lhe varlablllLy ln Lhe daLa as posslble) and each succeedlng
componenL ln Lurn has Lhe hlghesL varlance posslble under Lhe consLralnL LhaL lL
be orLhogonal Lo (uncorrelaLed wlLh) Lhe precedlng componenLs rlnclpal
componenLs are guaranLeed Lo be lndependenL only lf Lhe daLa seL ls [olnLly
normally dlsLrlbuLed CA ls senslLlve Lo Lhe relaLlve scallng of Lhe orlglnal
varlables uependlng on Lhe fleld of appllcaLlon lL ls also named Lhe dlscreLe
karhunenLoeve Lransform (kL1) Lhe PoLelllng Lransform or proper orLhogonal
decomposlLlon (Cu)

CA was lnvenLed ln 1901 by karl earson1 now lL ls mosLly used as a Lool ln
exploraLory daLa analysls and for maklng predlcLlve models CA can be done by
elgenvalue decomposlLlon of a daLa covarlance maLrlx or slngular value
decomposlLlon of a daLa maLrlx usually afLer mean cenLerlng Lhe daLa for each
aLLrlbuLe 1he resulLs of a CA are usually dlscussed ln Lerms of componenL scores
(Lhe Lransformed varlable values correspondlng Lo a parLlcular case ln Lhe daLa)
and loadlngs (Lhe welghL by whlch each sLandarlzed orlglnal varlable should be
mulLlplled Lo geL Lhe componenL score) (Shaw 2003)

CA ls Lhe slmplesL of Lhe Lrue elgenvecLorbased mulLlvarlaLe analyses CfLen lLs
operaLlon can be LhoughL of as reveallng Lhe lnLernal sLrucLure of Lhe daLa ln a
way whlch besL explalns Lhe varlance ln Lhe daLa lf a mulLlvarlaLe daLaseL ls
vlsuallsed as a seL of coordlnaLes ln a hlghdlmenslonal daLa space (1 axls per
varlable) CA can supply Lhe user wlLh a lowerdlmenslonal plcLure a shadow
of Lhls ob[ecL when vlewed from lLs (ln some sense) mosL lnformaLlve vlewpolnL
1hls ls done by uslng only Lhe flrsL few prlnclpal componenLs so LhaL Lhe
dlmenslonallLy of Lhe Lransformed daLa ls reduced

CA ls closely relaLed Lo facLor analysls lndeed some sLaLlsLlcal packages (such as
SLaLa) dellberaLely conflaLe Lhe Lwo Lechnlques 1rue facLor analysls makes
dlfferenL assumpLlons abouL Lhe underlylng sLrucLure and solves elgenvecLors of a
sllghLly dlfferenL maLrlx
lndependenL componenL analysls (lCA) ls a compuLaLlonal meLhod for separaLlng
a mulLlvarlaLe slgnal lnLo addlLlve subcomponenLs supposlng Lhe muLual sLaLlsLlcal
lndependence of Lhe nonCausslan source slgnals lL ls a speclal case of bllnd
source separaLlon

lnLroducLlon

When Lhe lndependence assumpLlon ls correcL bllnd lCA separaLlon of a mlxed
slgnal glves very good resulLs lL ls also used for slgnals LhaL are noL supposed Lo
be generaLed by a mlxlng for analysls purposes A slmple appllcaLlon of lCA ls Lhe
cockLall parLy problem" where Lhe underlylng speech slgnals are separaLed from
a sample daLa conslsLlng of people Lalklng slmulLaneously ln a room usually Lhe
problem ls slmpllfled by assumlng no Llme delays or echoes An lmporLanL noLe Lo
conslder ls LhaL lf n sources are presenL aL leasL n observaLlons (eg
mlcrophones) are needed Lo geL Lhe orlglnal slgnals 1hls consLlLuLes Lhe square (!
u where u ls Lhe lnpuL dlmenslon of Lhe daLa and ! ls Lhe dlmenslon of Lhe
model) CLher cases of underdeLermlned (! u) and overdeLermlned (! u) have
been lnvesLlgaLed

A hldden Markov model (PMM) ls a sLaLlsLlcal Markov model ln whlch Lhe sysLem
belng modeled ls assumed Lo be a Markov process wlLh unobserved (hldden)
sLaLes An PMM can be consldered as Lhe slmplesL dynamlc 8ayeslan neLwork

ln a regular Markov model Lhe sLaLe ls dlrecLly vlslble Lo Lhe observer and
Lherefore Lhe sLaLe LranslLlon probablllLles are Lhe only parameLers ln a hldden
Markov model Lhe sLaLe ls noL dlrecLly vlslble buL ouLpuL dependenL on Lhe
sLaLe ls vlslble Lach sLaLe has a probablllLy dlsLrlbuLlon over Lhe posslble ouLpuL
Lokens 1herefore Lhe sequence of Lokens generaLed by an PMM glves some
lnformaLlon abouL Lhe sequence of sLaLes noLe LhaL Lhe ad[ecLlve hldden refers
Lo Lhe sLaLe sequence Lhrough whlch Lhe model passes noL Lo Lhe parameLers of
Lhe model even lf Lhe model parameLers are known exacLly Lhe model ls sLlll
hldden

Pldden Markov models are especlally known for Lhelr appllcaLlon ln Lemporal
paLLern recognlLlon such as speech handwrlLlng gesLure recognlLlon parLof
speech Lagglng muslcal score followlng parLlal dlscharges and blolnformaLlcs

A hldden Markov model can be consldered a generallzaLlon of a mlxLure model
where Lhe hldden varlables (or laLenL varlables) whlch conLrol Lhe mlxLure
componenL Lo be selecLed for each observaLlon are relaLed Lhrough a Markov
process raLher Lhan lndependenL of each oLher

AnlsoLroplc dlffuslon


lrom Wlklpedla Lhe free encyclopedla




ln lmage processlng and compuLer vlslon anlsoLroplc dlffuslon also called
eronaMallk dlffuslon ls a Lechnlque almlng aL reduclng lmage nolse wlLhouL
removlng slgnlflcanL parLs of Lhe lmage conLenL Lyplcally edges llnes or oLher
deLalls LhaL are lmporLanL for Lhe lnLerpreLaLlon of Lhe lmage123 AnlsoLroplc
dlffuslon resembles Lhe process LhaL creaLes a scalespace where an lmage
generaLes a parameLerlzed famlly of successlvely more and more blurred lmages
based on a dlffuslon process Lach of Lhe resulLlng lmages ln Lhls famlly are glven
as a convoluLlon beLween Lhe lmage and a 2u lsoLroplc Causslan fllLer where Lhe
wldLh of Lhe fllLer lncreases wlLh Lhe parameLer 1hls dlffuslon process ls a llnear
and spacelnvarlanL LransformaLlon of Lhe orlglnal lmage AnlsoLroplc dlffuslon ls
a generallzaLlon of Lhls dlffuslon process lL produces a famlly of parameLerlzed
lmages buL each resulLlng lmage ls a comblnaLlon beLween Lhe orlglnal lmage and
a fllLer LhaL depends on Lhe local conLenL of Lhe orlglnal lmage As a consequence
anlsoLroplc dlffuslon ls a nonllnear and spacevarlanL LransformaLlon of Lhe
orlglnal lmage

ln lLs orlglnal formulaLlon presenLed by erona and Mallk ln 19871 Lhe space
varlanL fllLer ls ln facL lsoLroplc buL depends on Lhe lmage conLenL such LhaL lL
approxlmaLes an lmpulse funcLlon close Lo edges and oLher sLrucLures LhaL should
be preserved ln Lhe lmage over Lhe dlfferenL levels of Lhe resulLlng scalespace
1hls formulaLlon was referred Lo as anlsoLroplc dlffuslon by erona and Mallk
even Lhough Lhe locally adapLed fllLer ls lsoLroplc buL lL has also been referred Lo
as lnhomogeneous and nonllnear dlffuslon4 or eronaMallk dlffuslon3 by
oLher auLhors A more general formulaLlon allows Lhe locally adapLed fllLer Lo be
Lruly anlsoLroplc close Lo llnear sLrucLures such as edges or llnes lL has an
orlenLaLlon glven by Lhe sLrucLure such LhaL lL ls elongaLed along Lhe sLrucLure and
narrow across As a consequence Lhe resulLlng lmages preserve llnear sLrucLures
whlle aL Lhe same Llme smooLhlng ls made along Lhese sLrucLures 8oLh Lhese
cases can be descrlbed by a generallzaLlon of Lhe usual dlffuslon equaLlon where
Lhe dlffuslon coefflclenL lnsLead of belng a consLanL scalar ls a funcLlon of lmage
poslLlon and assumes a maLrlx (or Lensor) value

AlLhough Lhe resulLlng famlly of lmages can be descrlbed as a comblnaLlon
beLween Lhe orlglnal lmage and spacevarlanL fllLers Lhe locally adapLed fllLer and
lLs comblnaLlon wlLh Lhe lmage do noL have Lo be reallzed ln pracLlce AnlsoLroplc
dlffuslon ls normally lmplemenLed by means of an approxlmaLlon of Lhe
generallzed dlffuslon equaLlon each new lmage ln Lhe famlly ls compuLed by
applylng Lhls equaLlon Lo Lhe prevlous lmage ConsequenLly anlsoLroplc dlffuslon
ls an lLeraLlve process where a relaLlvely slmple seL of compuLaLlon are used Lo
compuLe each successlve lmage ln Lhe famlly and Lhls process ls conLlnued unLll a
sufflclenL degree of smooLhlng ls obLalned


lmage LnhancemenL
lmage enhancemenL ls Lhe lmprovemenL of dlglLal lmage quallLy (wanLed eg for
vlsual lnspecLlon or for machlne analysls) wlLhouL knowledge abouL Lhe source of
degradaLlon lf Lhe source of degradaLlon ls known one calls Lhe process lmage
resLoraLlon 8oLh are lconlcal processes vlz lnpuL and ouLpuL are lmages
Many dlfferenL ofLen elemenLary and heurlsLlc meLhods are used Lo lmprove
lmages ln some sense 1he problem ls of course noL well deflned as Lhere ls no
ob[ecLlve measure for lmage quallLy Pere we dlscuss a few reclpes LhaL have
shown Lo be useful boLh for Lhe human observer and/or for machlne recognlLlon
1hese meLhods are very problemorlenLed a meLhod LhaL works flne ln one case
may be compleLely lnadequaLe for anoLher problem

AparL from geomeLrlcal LransformaLlons some prellmlnary greylevel ad[usLmenLs
may be lndlcaLed Lo Lake lnLo accounL lmperfecLlons ln Lhe acqulslLlon sysLem
1hls can be done plxel by plxel callbraLlng wlLh Lhe ouLpuL of an lmage wlLh
consLanL brlghLness lrequenLly spacelnvarlanL greyvalue LransformaLlons are
also done for conLrasL sLreLchlng range compresslon eLc 1he crlLlcal dlsLrlbuLlon
ls Lhe relaLlve frequency of each greyvalue Lhe greyvalue hlsLogram

lmage resLoraLlon

lmage resLoraLlon refers Lo Lhe recovery of an orlglnal slgnal from degraded
observaLlons

lmage resLoraLlon ls dlfferenL from lmage enhancemenL ln LhaL Lhe laLLer ls
deslgned Lo emphaslze feaLures of Lhe lmage LhaL make Lhe lmage more pleaslng
Lo Lhe observer buL noL necessarlly Lo produce reallsLlc daLa from a sclenLlflc
polnL of vlew lmage enhancemenL Lechnlques (llke conLrasL sLreLchlng or de
blurrlng by a nearesL nelghbor procedure) provlded by lmaglng packages use no
a prlorl model of Lhe process LhaL creaLed Lhe lmage

WlLh lmage enhancemenL nolse can be effecLlvely be removed by sacrlflclng some
resoluLlon buL Lhls ls noL accepLable ln many appllcaLlons ln a lluorescence
Mlcroscope resoluLlon ln Lhe zdlrecLlon ls bad as lL ls More advanced lmage
processlng Lechnlques musL be applled Lo recover Lhe ob[ecL



lmage compresslon




Lmall



rlnL
AAAAAA
Llnkedln
lacebook
1wlLLer
Share 1hls
8eprlnLs


lmage compresslon ls mlnlmlzlng Lhe slze ln byLes of a graphlcs flle wlLhouL
degradlng Lhe quallLy of Lhe lmage Lo an unaccepLable level 1he reducLlon ln flle
slze allows more lmages Lo be sLored ln a glven amounL of dlsk or memory space
lL also reduces Lhe Llme requlred for lmages Lo be senL over Lhe lnLerneL or
downloaded from Web pages

1here are several dlfferenL ways ln whlch lmage flles can be compressed lor
lnLerneL use Lhe Lwo mosL common compressed graphlc lmage formaLs are Lhe
!LC formaL and Lhe Cll formaL 1he !LC meLhod ls more ofLen used for
phoLographs whlle Lhe Cll meLhod ls commonly used for llne arL and oLher
lmages ln whlch geomeLrlc shapes are relaLlvely slmple


L|terature rev|ew

Image rocess|ng and Ana|ys|s Var|at|ona| DL Wave|et and Stochast|c
Methods
1ony I Chan and I|anhong (Iack|e) Shen
Soc|ety for Industr|a| and App||ed Mathemat|cs (SIAM)
h||ade|ph|a A 200S xx|| + 400 pages IS8N 089871S89k

kece|ved May 24 2006 Accepted Iune 9 2006
lmage processlng ls an lmporLanL componenL of modern Lechnologles because
human depends so much on Lhe vlsual lnformaLlon Lhan oLher creaLures lmage ls
beLLer Lhan any oLher lnformaLlon form for us Lo percelve Among our
lnformaLlon abouL Lhe world 99 ls percelved wlLh our eyes 1 lmage
processlng has LradlLlonally been an area ln Lhe englneerlng communlLy 1he baslc
Lools are lourler analysls wlLh a long hlsLory and waveleL analysls whlch has
become popular ln Lhe englneerlng communlLy slnce 1980s ln Lhe pasL few
decades several advanced maLhemaLlcal approaches have been lnLroduced lnLo
Lhls fleld namely Lhe varlaLlonal calculus parLlal dlfferenLlal equaLlons (uL) and
sLochasLlc approaches (or sLaLlsLlcal meLhods) and Lhey have become lmporLanL
Lools for LheoreLlcal lmage processlng lrom Lhe cover 1hls book brldges
conLemporary maLhemaLlcs wlLh sLaLeofLhearL meLhodologles ln modern lmage
processlng whlle organlzlng Lhe vasL conLemporary llLeraLure lnLo a coherenL and
loglcal sLrucLure

ln Lhls book Lhe auLhors presenL an lnLegraLed Lheory for lmage processlng
Lhrough Lhe aforemenLloned approaches lrom Lhe preface 1hese dlverslfled
approaches are apparenLly dlsLlncL buL ln facL lnLrlnslcally connecLed Lach
meLhod excels from cerLaln lnLeresLlng angle or levels of approxlmaLlons buL ls
also lnevlLably sub[ecL Lo lLs llmlLaLlons and appllcablllLles Cn Lhe oLher hand aL
some deeper levels Lhey share common grounds and rooLs from whlch more
efflclenL hybrld Lools or meLhods can be developed 1hls hlghllghLs Lhe necesslLy
of lnLegraLlng Lhls dlverslLy of approaches

1here ls an lnLroducLory preface and 7 chapLers many lllusLraLlon flgures an
lndex and an exLended blbllography of 330 enLrles

1he 1sL chapLer lnLroducLlon beglns wlLh an lnLroducLlon of lmaglng sclence
and lmage processlng problems covered ln Lhe book lmage enhancemenL lmage
denolslng lmage deblurrlng lmage lnpalnLlng and lmage segmenLaLlon lL
presenLs a meLhodologlcal overvlew of Lhe lourler analysls maLhemaLlcal
morphology waveleL Lheory sLochasLlc approach varlaLlonal meLhod and parLlal
dlfferenLlal equaLlons 1he auLhors conclude LhaL Lhose varlous approaches are
lnLrlnslcally lnLerconnecLed 1here ls one secLlon abouL Lhe organlzaLlon of Lhe
book and one on Lhe usage of Lhe book for readers of dlfferenL backgrounds

1he 2nd chapLer Some Modern lmage Analysls 1ools revlews maLhemaLlcal
Lheorles for Lhe geomeLry of curves/surfaces funcLlons of bounded varlaLlons
(8v) and 8ayeslan sLaLlsLlcal lnference llnear/nonllnear Lechnlques of
fllLerlng/dlffuslon and waveleLs/mulLlresoluLlon analysls 1hls chapLer can also
serve as a reference source for readers from oLher dlsclpllnes oLher Lhan
maLhemaLlcs

1he 3rd chapLer lmage Modellng and 8epresenLaLlon dlscusses flve
maLhemaLlcal represenLaLlons of lmages deLermlnlsLlc models as funcLlons of
varlous smooLhness sLochasLlc models wlLh random flelds mulLlscale models
based on Lhe waveleL Lheory levelseL models by uLs and MumfordShah
models from Lhe varlaLlonal calculus 1hese represenLaLlons llnk Lhe maLhemaLlcal
Lheorles ln Lhe 2nd chapLer wlLh lmage processlng problems ln remalnlng
chapLers

1he 4Lh chapLer lmage rocesslng uenolslng flrsL revlews Lhe physlcal orlgln
and sLochasLlc models of lmage nolse lL flrsL recalls Lhe classlcal denolslng
Lechnlques such as llnear lowpass fllLers and Wlener fllLers 1he followlng
modern denolslng Lechnlques are Lhen presenLed Lhe semlnal waveleL shrlnkage
Lechnlque varlaLlonal denolslng meLhods based on 1v and 8v models and
algorlLhms from llnear/nonllnear dlffuslon and scale space Lheory 1here ls an
appllcaLlon Lo salLandpepper nolse removal 1he exLenslon Lo mulLlchannel 1v
denolslng ls provlded aL Lhe end of Lhe chapLer

As ln Lhe 4Lh chapLer Lhe 3Lh chapLer lmage rocesslng ueblurrlng also flrsL
revlews Lhe physlcal background and sLochasLlc models of lmage blur 1he lll
posedness of Lhe deblurrlng problem and regularlzaLlon Lechnlques are Lhen
brlefly addressed Lo whlch Lhere are many books from dlfferenL perspecLlves
devoLed lmage deblurrlng meLhods ln Lhls chapLer lnclude Lhe classlcal Wlener
fllLraLlon Lechnlque and varlaLlonal meLhods for known Sls 8llnd deblurrlng for
unknown Sls wlLh parameLrlc and nonparameLrlc Sls are also presenLed

1he 6Lh chapLer lmage rocesslng lnpalnLlng ls Lo lnvesLlgaLe Lhe lmage
lnpalnLlng problem from Lhe perspecLlve of lnLerpolaLlon Lheory 1he word
lnpalnLlng ls an arLlsLlc synonym for lmage lnLerpolaLlon p243 varlous
meLhods are sLudled lncludlng classlcal polynomlal LrlgonomeLrlc spllne and
radlal base funcLlon lnLerpolaLlons and recenL Lechnlques wlLh 1v 8adon
measure MunfordShahLuler model waveleL uLs and Clbbs/Markov flelds
Several appllcaLlons of lnpalnLlng Lechnlques are also dlscussed

1he 7Lh chapLer lmage rocesslng SegmenLaLlon presenLs recenL lmage
segmenLaLlon meLhods lncludlng Lhe acLlve conLour Ceman and Cemans Clbbs
mlxLure model MumfordShahs free boundary model and mulLlchannel loglcal
segmenLaLlon Lechnlque unllke oLher chapLers Lhls chapLer does noL provlde a
deLalled background lnLroducLlon on lmage segmenLaLlon 8eaders can flnd
relevanL lnformaLlon ln 13

1he book presenLs four powerful maLhemaLlcal Lools for lmage processlng and
analysls varlaLlonal calculus uL waveleL and sLochasLlc meLhods 1he auLhors
have subsLanLlally conLrlbuLed Lo Lhe developmenL of Lhe Lheorles and
Lechnlques 1he book ls well developed ln depLh and Lo some exLenL self
conLalned Cenerally speaklng Lhls book ls ln hlgh quallLy and should be of general
lnLeresLs Lo Lhe lmaglng processlng communlLy and relevanL flelds Lhough Lhe
underlylng maLhemaLlcs may be a burden Lo some readers lor example some
parLs are qulLe maLhemaLlcal eg 71 on Lhe monolds of occluslve prelmages
1he algorlLhms ln Lhe book may noL be easy Lo lmplemenL because numerlcal
Lechnlques and programmlng Llps are noL easy for Lhe compleLe novlce even Lhe
auLhors have made a greaL and successful efforL ln Lhe algorlLhm presenLaLlons lL
would be very helpful lf Lhe auLhors can publlsh some demo codes on Lhe
lnLerneL neverLheless Lhe revlewer concurs wlLh Lhe followlng commenL by !ean
Mlchel Morel ln Lhe cover Lhls book ls A compleLe and very lnformaLlve revlew
of maLhemaLlcal advances ln lmage analysls ln Lhe pasL LwenLy years

lrom Lhe cover 1hls book ls wrlLLen for graduaLe sLudenLs and researchers ln
applled maLhemaLlcs compuLer sclence elecLrlcal englneerlng and oLher
dlsclpllnes who are lnLeresLed ln problems ln lmaglng and compuLer vlslon lL can
be used as a reference by sclenLlsLs wlLh speclflc Lasks ln lmage processlng as well
as by researchers wlLh a general lnLeresL ln flndlng ouL abouL Lhe laLesL advances
lnLeresLed readers are recommended Lo consulL relevanL books for lmage
processlng from Lhe maLhemaLlcal perspecLlve 317

keferences
18uss !C 1he lmage processlng handbook 2nd 8oca 8aLon C8C ress 1993 p
674 p
2Marr u vlslon a compuLaLlonal lnvesLlgaLlon lnLo Lhe human represenLaLlon
and processlng of vlsual lnformaLlon San lranclsco WP lreeman 1982 p xvll
397 p
3Morel !M Sollmlnl S rogress ln nonllnear dlfferenLlal equaLlons and Lhelr
appllcaLlons v 14 8osLon 8lrkhauser 1993 varlaLlonal meLhods ln lmage
segmenLaLlon wlLh seven lmage processlng experlmenLs p xvl 243 p
4AuberL C kornprobsL MaLhemaLlcal problems ln lmage processlng parLlal
dlfferenLlal equaLlons and Lhe calculus of varlaLlons new ?ork Sprlnger 2001 p
xxv 286 p
3ual Maso C 1omarelll l rogress ln nonllnear dlfferenLlal equaLlons and Lhelr
appllcaLlons v 31 8asel 8osLon 8lrkhauser 2002 varlaLlonal meLhods for
dlsconLlnuous sLrucLures lnLernaLlonal workshop aL vllla Lrba (Cernobblo) lLaly
!uly 2001 p x 183 p
6Paar 8omeny 8 LecLure noLes ln compuLer sclence 1232 new ?ork Sprlnger
1997 Scalespace Lheory for compuLer vlslon 1sL lnLernaLlonal Conference Scale
Space 97 uLrechL 1he neLherlands !uly 24 1997 proceedlngs p lx 364 p
7nlelsen M LecLure noLes ln compuLer sclence 1682 8erlln new ?ork
Sprlnger 1999 Scalespace Lheorles ln compuLer vlslon 2nd lnLernaLlonal
conference ScaleSpace99 Corfu Creece SepLember 2627 1999 proceedlngs
p xll 332 p
8kerckhove M LecLure noLes ln compuLer sclence 2106 new ?ork Sprlnger
2001 Scalespace and morphology ln compuLer vlslon 3rd lnLernaLlonal
Conference ScaleSpace 2001 vancouver Canada !uly 78 2001 proceedlngs p
xl 434 p
9Crlffln Lu Llllholm M LecLure noLes ln compuLer sclence 2693 8erlln new
?ork Sprlnger 2003 Scale space meLhods ln compuLer vlslon 4Lh lnLernaLlonal
conference Scale Space 2003 lsle of Skye uk !une 1012 2003 proceedlngs p
xlll 816 p
10klmmel 8 Sochen n WelckerL ! LecLure noLes ln compuLer sclence 3439
8erlln Sprlnger 2003 Scale space and uL meLhods ln compuLer vlslon 3Lh
lnLernaLlonal conference ScaleSpace 2003 Pofgelsmar Cermany Aprll 79
2003 proceedlngs p xl 634 p
11Csher S araglos n CeomeLrlc level seL meLhods ln lmaglng vlslon and
graphlcs new ?ork Sprlngerverlag 2003 p xxlx 313 p
12araglos n Chen ? laugeras C Pandbook of maLhemaLlcal models ln
compuLer vlslon new ?ork Sprlnger 2006 p xxxlll 603 p
138onse C na[man L uecenclere L CompuLaLlonal lmaglng and vlslon v 30
uordrechL Sprlnger 2003 MaLhemaLlcal morphology 40 years on proceedlngs
of Lhe 7Lh lnLernaLlonal Symposlum on MaLhemaLlcal Morphology Aprll 1820
2003 p xvll 490 p
14Sporrlng ! CompuLaLlonal lmaglng and vlslon v 8 uordrechL 8osLon kluwer
Academlc ubllshers 1997 Causslan scalespace Lheory p xx 263 p
13WelckerL ! LCMl Serles SLuLLgarL 1eubnerverlag 1998 AnlsoLroplc
dlffusslon ln lmage processlng p xll 170 p
16Wlnkler C AppllcaLlons of maLhemaLlcs 27 2nd 8erlln new ?ork
Sprlngerverlag 2003 lmage analysls random flelds and Markov chaln MonLe
Carlo meLhods a maLhemaLlcal lnLroducLlon p xvl 387 p
17araglos n LecLure noLes ln compuLer sclence 3732 8erlln new ?ork
Sprlnger 2003 varlaLlonal geomeLrlc and level seL meLhods ln compuLer vlslon
3rd lnLernaLlonal workshop vLSM 2003 8el[lng Chlna CcLober 16 2003
proceedlngs p xl 367 p