Digital Image Inpainting Using Patch Priority Based Method

8/12/2019 Digital Image Inpainting Using Patch Priority Based Method

1/12

International Journal of Advance Foundation and Research in Computer (IJAFRC)

Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

*i+ital Ima+e Inpaintin+ sin+ -atch -riorit .ased MethodGeeta K. Sarpate, Shanti K. Guru

PG Student, Department of Computer Engineering, D.Y.Patil College of Engineering,Assistant Professor, Department of Computer Engineering, D.Y.Patil College of Engineering,

Akurdi, Pune-4 !44, "aharashtra, #ndia

sarpategeeta$%gmail.&om

A . % / R A C /

In recent ear, di+ital ima+e inpaintin+ is an interestin+ ne0 research topic in ima+e processin+,

0hich can e used in man applications li2e computer +raphics, ima+e editin+, film

postproduction, ima+e restoration$ It can e used in special effects and the restoration of old

photo+raphs and dama+ed film, removal of superimposed te3t li2e dates, sutitles, or pulicit

and the removal of entire o4ects from the ima+e$ In ima+e inpaintin+, missin+ (tar+et) re+ions

0ere filled propa+atin+ structural and te3tural information of an ima+e in a visuall plausile

0a, also 2no0n as ima+e inpaintin+$ /hou+h this techniue is ver useful, it is still a challen+in+

prolem in computer +raphics and computer vision$ In this paper, an al+orithm is proposed for

removin+ tar+et o4ects from di+ital ima+es$ In addition, al+orithms 0as proposed to

snthesi6ethe structure 7 te3ture as 0ell as fill the hole that is left ehind inan undetectale form$

An attempt has een made to compute actual color values usin+ e3emplar ased snthesis and

patch priorit$ A numer of e3amples on removin+ occludin+ o4ects from real and snthetic

ima+es demonstrate the effectiveness of proposed al+orithm in terms of oth inpaintin+ ualit

and computational efficac$ All e3periments 0ere run on a !$5 896 -entium V 0ith ! 8. of

RAM,1:"8. or aove 9*$Inde3 /erms ; Ima+e inpaintin+, Al+orithm, /e3tural snthesis, %tructural snthesis, -atch

-riorit, Re+ion fillin+

I$ I&/R


2/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

gra( le/el of the image at that point. 8hen 0, ( and the amplitude /alues of are all finite, dis&rete

'uantities, *e &all the image a digital image. 9ote that a digital image is &omposed of a finite num)er ofelements, ea&h of *hi&h has a parti&ular lo&ation and /alue. hese elements are referred to as pi&ture

elements, image elements and pi0els. #mages usuall( get &orrupted )e&ause of loss in transmission,storage. 8e need to de/elop the algorithms *hi&h &an automati&all( fill that lost information+:,;.


3/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

algorithm *hi&h &an fill in a region di&tated )( the user.he end goal of digital image inpainting *ould )e

to ha/e the user to pro/ide the algorithm *ith the region to )e inpainted, *hi&h shall )e referred to asthe mask, and then to ha/e the &omputer automati&all( fill in this region in a manner su&h that the

resulting image looks natural and unaltered. "ost inpainting methods *ork as follo*s. @irst, the imageregions to )e inpainted are sele&ted, usuall( manuall(+B, !. 9e0t, &olor information is propagated

in*ard from the region a )oundar(, i.e., the kno*n image information is used to fill in the missing areas.

#n order to produ&e a per&eptuall( plausi)le re&onstru&tion, an inpainting te&hni'ue should attempt to&ontinue the isophotes =lines of e'ual gra( /alue> as smoothl( as possi)le inside the re&onstru&ted

region. he pro&ess of remo/ing o)7e&ts from images starts *ith mask out the undesired o)7e&t, makingthe area *here the o)7e&t pre/iousl( o&&upies a gap. hen the gap *ill )e filled using graphi&al

te&hni'ues su&h as inpainting. Among the graphi&al te&hni'ues that are used to fill the gap after o)7e&tremo/al, t*o most &ommonl( used are image inpainting and te0ture s(nthesis +.#n this paper, an

algorithm for automati& digital inpainting, )eing its main moti/ation to repli&ate the )asi& te&hni'ues

used )( professional restorators is introdu&ed. At this point, the onl( user intera&tion re'uired )( the

algorithm here introdu&ed is to mark the regions to )e inpainted. "oreo/er, sin&e the inpaintingalgorithm here presented &an )e used not 7ust to restore damaged photographs )ut also to remo/e

undesired o)7e&ts and *ritings on the image, the regions to )e inpainted must )e marked )( the user,

sin&e the( depend on hisher su)7e&ti/e sele&tion. ere *e are &on&erned on ho* to. @ill-in

V$ MI/=RA/R= R=VI=?he &on&ept of digital inpainting *as first initiated in the paper )( Fertalmio et al. +. Smart digital

inpainting models, te&hni'ues, and algorithms ha/e man( appli&ations in image interpolation, photo

restoration, 6ooming and super-resolution primal-sket&h )ased per&eptual image &ompression and

&oding, and the error disguise of =*ireless> image transmission, et&. #mage inpainting is an image

restoration pro)lem, in *hi&h image models pla( a &riti&al role, as demonstrated )( Chan, Kang, and

Shens + re&ent inpainting s&hemes )ased on the otal Hariation and elasti&all( image models. here are

/arious approa&hes to image pro&essing, su&h as sto&hasti& modeling, *a/elets and partial differential

e'uation =PDE> approa&hes. he last t*o de&ades or so PDE methods ha/e )een /er( popular. #mage

pro&essing in&ludes a /ast num)er of fields. #mage inpainting is a method of modif(ing images in su&h a

manner that one &annot dete&t the modifi&ation to the image. A t(pi&al appli&ation of inpainting is therestoration of old and de&a(ing paintings. #n this s&enario, &ra&ks or other defe&ts ha/e appeared in the

*ork and one *ishes to restore the pie&e so that there are no /isi)le defe&ts. An e0ample of *hen image

inpainting is needed is *hen one *ishes to remo/e an undesired o)7e&t or indi/idual from an

image.@i0ing images using inpainting has a long histor(. "ost nota)l(, during the renaissan&e, man(

medie/al art*orks had )een )rought Iup to dateI. "issing or damaged parts in the paintings *ere

re&onstru&ted in a *a( that the( are not dete&ta)le from human e(es. Stru&tures and te0tures around the

gap *ere &arefull( e0tended into the missing area. he results *ould look natural enough that o)ser/ers

*ithout prior kno*ledge of the original image *ill not noti&e the gaps+. algorithm imitates the

traditional inpainting pro&esses, su&h as determine the area to )e &orre&ted, e0am the )oundar( of the

region to )e filled, and &ontinuing lines of similar &olor. #n this /ie*, Kokaram et al.+5 used motionestimation and autoregressi/e models to interpolate losses in films from ad7a&ent frames. he )asi& idea

is to &op( into the gap the right pi0els from neigh)oring frames. he te&hni'ue &annot )e applied to still

images or to films *here the regions to )e inpainted span man( frames. #n addition, irani and


4/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

otsuka+4. &om)ined fre'uen&( and spatial domain information in order to fill a gi/en region *ith a

sele&ted te0ture. his is a /er( simple te&hni'ue that produ&es in&redi)le good results. , and re'uires

the user to sele&t the te0ture to )e &opied into the region to )e inpainted. #n this perspe&ti/e, "asnou and

"orel +: e0tended these ideas, presenting a /er( inspiring general /ariational formulation for

diso&&lusion and a parti&ular pra&ti&al algorithm =not entirel( )ased on PDEs> implementing some of the

ideas in this formulation. he algorithm performs inpainting )( 7oining *ith geodesi& &ur/es the points of

the isophotes =lines of e'ual gra( /alues> arri/ing at the )oundar( of the region to )e inpainted. #n

Fertalmio + and Fertalmio +; the image smoothness information, estimated )( the image Japla&ian, is

propagated along the isophotes dire&tions, estimated )( the image gradient rotated B! degrees.

he otal Hariational =H> model +? uses an Euler-Jagrange e'uation &oupled *ith an isotropi& diffusion

to maintain the isophotes dire&tions. he Cur/ature-Dri/en Diffusion =CCD> model enhan&es the H

method to dri/e diffusion along the isophotes dire&tions and thus allo*s inpainting of thi&ker regions +.herefore, the pat&h-)ased approa&h of Criminisi et al. +$ is more appropriate. he sour&e pat&h =the

pat&h from the kno*n area> &an )e rotated, s&aled, flipped, or its intensit( &an )e ad7usted to )ettermat&h the target pat&h =the pat&h )eing re&onstru&ted>. here is a large amount of literature using a

&ertain degree of in/arian&e in mat&hing pro)lems +B. @or image re&onstru&tion, these te&hni'ues ha/e)een adopted in Drori et al.+!. #n re&ent literature, Anupam et al.+ presented an algorithm that &an

remo/e o)7e&ts from the image in a *a( that it seems reasona)le to the human e(e. #t &an restore old

photographs =e.g. remo/al of S&rat&hes>. ong)en and Lian + proposed a no/el pat&h propagation

)ased #npainting algorithm for s&rat&h or te0t remo/al, o)7e&t remo/al and missing )lo&k &ompletion.his stud( re/ealed that the proposed e0amplar-)ased pat&h propagation algorithm &an )etter infer the

stru&tures and te0tures of the missing region, and produ&e sharp inpainting results &onsistent *ith the

surrounding te0tures. Shutao and hao +5 presented a no/el inpainting method )ased on automati&

salient stru&ture &ompletion. Msing in&omplete salient stru&ture e0tra&tion and &ompletion, method&ompletes the stru&ture of missing region reasona)l(. he &ompleted salient stru&tures di/ide the target

area into se/eral su)-regions. hen, te0ture propagation is used to s(nthesi6e the te0ture information*ith samples from &orresponding ad7a&ent su)-regions. his redu&es the running-time and offers morepre&ise te0ture information. Soma(eh et al. +4 presented a modified e0amplar-)ased inpainting method

in the frame*ork of pat&h sparsit(. #n the e0amplar-)ased algorithms, the unkno*n )lo&ks of target

region are inpainted )( the most similar )lo&ks e0tra&ted from the sour&e region, *ith the a/aila)le

information. Defining a priorit( term to de&ide the filling order of missing pi0els ensures the &onne&ti/it(of o)7e&t )oundaries. #n the e0emplar-)ased pat&h sparsit( approa&hes, a sparse representation of

missing pi0els *as &onsidered to define a ne* priorit( term. "oreo/er, Ji0in and Chen +: presented animpro/ed method for e0emplar-)ased image inpainting. F( &onsidering the isophote &ur/ature as a ne*

ingredient to &ompute pat&h priorit( and mat&hing &ost fun&tion, this method performs *ell *hen thetarget region &ontains linear stru&ture *ith signifi&ant &hanges in &ur/ature. Feside this, Dang et al. +;

introdu&ed a no/el unsuper/ised image &ompletion frame*ork using a modified e0emplar-)ased methodin &on7un&tion *ith a p(ramidal representation of an image. A top-do*n iterati/e &ompletion is

performed graduall( *ith multi-resolution pat&hes and a *indo*-)ased priorit(. he proposed approa&h

is /erified on different natural images. 8ang et al.+? proposed an impro/ed e0emplar-)ased image

inpainting method for remo/ing o)7e&ts in digital images. he( introdu&ed a regulari6ed fa&tor, *hi&had7usts the &ur/e of the pat&h priorit( fun&tion, in &omputing the filling order. hese impro/ementsa&tuall( make the #npainting more ro)ust to images *ith the large remo/al regions. Feside this, 9eelima

and Arul/an+$. proposed a )est algorithm in *hi&h the &onfiden&e in the s(nthesi6ed pi0el /alues ispropagated in a manner similar to the propagation of information in inpainting. he a&tual &olour /alues

are &omputed using e0emplar-)ased s(nthesis.

#n this paper, an algorithm is proposed for remo/ing large o)7e&ts from digital images. he &hallenge is to

fill in the hole that is left )ehind in a /isuall( plausi)le *a(. An algorithm is de/ised in *hi&h the&onfiden&e in the s(nthesi6ed pi0el /alues is propagated in a manner similar to the propagation of

information in inpainting. he a&tual &olour /alues are &omputed using e0emplar-)ased s(nthesis.#n

addition, an algorithm is proposed for remo/ing o)7e&ts from digital images and repla&ing them *ith


5/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

/isuall( plausi)le )a&kgrounds. he algorithm effe&ti/el( imagines ne* &olour /alues for the target

region in a *a( that looks IrationalI to the e(e. #n pre/ious *ork, se/eral resear&hers ha/e &onsideredte0ture s(nthesis as a *a( to fill large image regions *ith IpureI te0tures - repetiti/e t*o dimensional

te0tural patterns *ith moderate sto&hasti&it( this has )een addressed )( t*o &lasses of algorithms =i>e0ture s(nthesis and =ii> #npainting. he pre/ious *ork *ell for Ite0turesI - repeating t*o dimensional

patterns *ith some sto&hasti&it( the se&ond fo&us on linear Istru&turesI *hi&h &an )e supposed of as one

dimensional patterns, su&h as lines and o)7e&t &ontours. An effi&ient algorithm that &om)ines thead/antage of these t*o methods is intended. 9e0t se&tion e0plains te0ture s(nthesis and further se&tion

illustrates the region filling )( proposed algorithm.

VII$ C@%I%9o* *e &ompare the different algorithm use for digital image inpainting. he &omparison )et*eendifferent inpainting algorithm, ad/antage, and disad/antage is listed in a)le .

/ale1$Comparative Analsis of Inpaintin+ Al+orithm

VIII$ %@%/=M ARC9I/=C/R=S(stem Ar&hite&ture are sho*n in @igure 5. S(stems design is the pro&ess of defining the ar&hite&ture,interfa&es, modules, &omponents, modules, and data for a s(stem to satisf( spe&ified re'uirements.

A. Input Ima+e; #nput image is a photograph, natural image and s(ntheti& image for pro&essing that&ontain &olored images of different si6e et&. that is gi/en to image inpainting.

F. Mas2 Ima+e;Nemo/e the un*anted o)7e&t either single or multiple if re'uired from the originalimage through paint &alled mask image or target *hi&h is to )e inpainted.

Al+orithm Advanta+e *isadvanta+ePDE Fased #mage

#npainting Algorithm

#t *orks *ell in images that

are relati/el( smooth O do

not

&ontain too mu&h noise orte0ture.

he( are una)le to fill in

regions that are highl(

te0tured or &ontain too

mu&h noise.

e0ture S(nthesis

Algorithms

hese s(nthesis )ased

te&hni'ues perform*ell onl(

for a sele&ted set of images.

hese algorithms ha/e

diffi&ult( in

handlingnatural imagesas the( are &omposed of

stru&tures in the form of

edges.

An E0emplar Fased#npainting Algorithm

E0emplar-)ased inpainting*orks *ell in &ases of regular

te0tures.

his algorithm generatethe Istair&ase effe&t Iin

image inpainting.

E0emplar Fased

Pat&h Sparsit(

he( &an remo/e large

o)7e&ts from images

a&&ording to the definedpat&h priorit( /alue assigned

to the pi0el.

his algorithm does not

&al&ulate *ell pat&h

priorit( if te0ture is&ompli&ated.

An Effe&ti/e E0emplar

Fased #mage #npainting

Algorithm

#t &an propagate stru&ture

informationand te0ture

information.

his algorithm does not

perform *ell if

&ompli&ated stru&ture Ote0ture in stillphotographs.


6/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

C. Incomplete %alient %tructure Completion;Conne&ting in&omplete salient stru&tures is the ke( too)tain &redita)le inpainting results. o*e/er, simple &onne&tion or e0tension of the in&ompletesalient stru&tures results. te0ture features are used to determine the similarit( of in&omplete

salient stru&tures.

D. /e3ture -ropa+ation of Completed %alient %tructures; After stru&ture &ompletion, *epropagate te0ture information into the target region through pat&h )ased inpainting method =thepat&h si6e is B0B pi0els>.

Fi+ure '$ %stem Architecture

I$ %@%/=M IM->=M=&/A/I, &reate large non-repetiti/e )a&kground images and e0pand small pi&tures. #n

image pro&essing, e/er( digital image &omposed of repeated elements is &alled a Ite0ture.I e0tures(nthesis algorithms are intended to &reate an output image as similar as possi)le to the sample and thesi6e gi/en )( the user E0emplar )ased te0ture s(nthesis is suffi&ient for propagation in linear image

stru&tures as *ell and no separate s(nthesis me&hanism is re'uired. @igure 4 =a,),&,d> sho*s theE0emplar )ased te0ture s(nthesis)lank line )efore, and one after.

a> 8e *ant to s(nthesi6e area delimited )( the pat&h Tp &entered on point p RQ&> he most likel( &andidate mat&hes for Tp lie along the )oundar( )et*een the t*o te0tures in the

sour&e regionq =d>he )est mat&hing pat&h in the &andidates set has )een &opied into the

position o&&upied )( Tp, thus o)taining partial filling of Q. he target region Q has no* shrank

and its front has assumed a different shape.

Fi+ure #$ (a,,c,d) =3emplar ased te3ture snthesis


7/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

@igure 4 re/eals this point. @or ease of &omparison, *e used similar and &ommon notation as used ininpainting literature. he region to )e filled is sho*n )( Q, and its &ontour is sho*n )( RQ. he &ontour

mo/es in*ard as the algorithm applies. he sour&e region remains fi0ed throughout. Single iterationof the algorithm sho*s ho* stru&ture and te0ture are progressed )( e0emplar )ased s(nthesis.

Consider the s'uare template TpQ &entered at point p =figure 4)>, is to )e filled. he )est suita)lemat&h sample from sour&e region &omes from the pat&h

q Q. #n the e0ample in fig 4 ) *e find that if

Tp lies on &ontinuation of an image edge, the most likel( )est mat&hes *ill lie along the same edge. Allthis is needed to propagate the isophote in*ards is a simple transfer of the pattern from the )est mat&h

sour&e pat&h =figure 4d>.

.$ Re+ion Fillin+@irst the user sele&ts target region to )e remo/ed and filled. 9e0t, as *ith all e0emplar )ased te0tures(nthesis. he si6e of template *indo* must spe&if( in the sour&e region. is &omputed. P=p> is defined as the produ&t of t*o terms a &onfiden&e termC=p>, and a data term D=p>

P=p> U C=p>.D=p>, =>

( )(p)

| |

p

p

q C qC

=

,| .n |

( ) pIp

D p

= =>

8here, VTpV is the area of p, W is normali6ation fa&tor =e.g. W U :: for a t(pi&al gre(-le/el image>, npis a

unit /e&tor orthogonal to the front RQ in the point p andp

is an isophote /e&tor. D=p> en&ourages linear

stru&ture to )e s(nthesi6ed first and thus propagates se&urel( into the target region, C=p> illustrates the

amount of the relia)le information surrounding the pi0el p and is initiated to )e C=p> U !, p , andC=p> U , p .

he pat&h p *ith the highest priorit( is found to )e filled in *ith the information e0tra&ted from the

sour&e region .A glo)al sear&h is &arried out on the *hole image to find a pat&hq that has the most

similarit( *ith p . @ormall(,

argmin ( , )p q p qd = =5>

8here the distan&e ( , )p qd )et*een t*o generi& pat&hes a and b is simpl( defined as the sum of

s'uared differen&e =SSD> of the alread( filled pi0els in the t*o pat&hes.. he /alue of ea&h of pi0el to )e filled in, | pp p is &opied from its &orresponding position inside

q as follo*s


8/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

he &onfiden&e term C=p> is updated in the area delimited )( p as follo*s

(q) (p) q pC C = =4>

$ MA/9=MA/ICA> M. Jet S )e a s(stem that des&ri)es a digital image inpainting )( te0ture s(nthesis and region filling.

his s(stem allo*s a user to remo/e un*anted o)7e&t and to restore the damage image from

image fill. he images data)ase is needed to perform image inpainting.

S U X..Z

. #dentif( the Sour&e region.S U X[,.Z

[U Sour&e #mage.

5. #dentif( the target region.S U X[,2,..Z

2U arget region.

4. #dentif( the &ontour or )oundar( region.S U X[,2,R2,Z

R2 U Foundar( region.

:. #dentif( the pat&h &enter.S U X[,2,R2,Tp,.Z

p\R2

Tp U Pat&h pointer is element of )oundar( region.

;. #dentif( most likel( pat&h p along *ith )oundar( )et*een t*o te0ture in sour&eNegion [ e.g. T'S U X[,2,R2,Tp,T',.Z

?. #dentif( the )est mat&hing pat&h in &andidate set &opied into position o&&upied )( TpU T'and fill the target region.

S U X[,2,R2,Tp,T',Tp,..Z

$. Compute pat&h priorit( P=p>S U X[,2,R2,Tp,T',Tp,P=p>,..Z

P=p> U C=p>] D=p>C=p>U Confiden&e term, D=p>U Data term.


9/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

B. Confiden&e term C =p> measures the amount of relia)le information in the pat&hTpC=p> U ! for all p \2

C=p> U for all p \

D=p> U V=isophotes /e&tor> ] =unit normal /e&tor> V =normali6ation fa&tor>

!.#dentif( the pat&h *ith highest priorit( Tp^S U X[,2,R2,Tp,T',Tp,P=p>,Tp^,..Z

Tp^U a/erage minimum sum of s'uare differen&e )et*een t*o pat&h.

.Pla&e /alue of the pat&h *ith highest priorit( to fill in T'^S U =[,2,R2,Tp,T',Tp,P=p>,Tp^,T'^,.>

.Mpdate &onfiden&e term C=p>S U X[,2,R2s,Tp,T',Tp,P=p>,Tp^,T'^, C=p>Z

I$ A>8


10/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

Tp^U arg ma0 p\_2P=p>

%tep 5 @ind the e0emplar '^\ that minimi6es d=Tp^, T'^>

%tep : Cop( the kno* pi0els of pat&h Tp^to the &orresponding unkno*n pi0els of pat&h T'^and updatethe &onfiden&e /alue.C=p>U for All p V p\Tp^2

%tep B Nepeat a)o/e steps until all missing pi0els are filled &ompletel(.

II$ =-=RIM=&/A> R=%>/%8e implemented our approa&h and &ompared *ith the &on/entional approa&h for se/eral different

images ranging from purel( s(ntheti& images to full-&olor images in&luding &omple0 te0tures using

"AJAF instru&tions. #n all of the implementation, the pat&h si6e *as set to )e greater than the largest

pi0el or the thi&kest stru&ture =e.g. edges> in the sour&e region.#n figure ; the target o)7e&t *as remo/ed*ith plausi)le )a&kground and almost *ith nearest &olor. Similar results are sho*n in figure ?.

"oreo/er, the sour&e region has )een set to )e I = . All e0periments *ere run on a .: G Pentium

H *ith GF of NA".

Fi+ure :$ a)


11/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

III$ C/8e ha/e &ompared results of the restoration of an image. @igure $& sho*s the results of filling )(

proposed algorithm in *hi&h the hori6on is &orre&tl( re&onstru&ted as a straight line. =&>

Fi+ure $a) , pp. :!B- :B, BB:.

+4 A.9. irani and .otsuka, bDual Domain #ntera&ti/e #mage Nestoration Fasi& Algorithmc, #npro&eedings of #C#P BB;,pp.?B?-$!!,BB;.

+: S. "asnou and L."."orel, bJe/el Jines )ased diso&&lusionc,#n pro&eedings of #C#P B$,pp.:B-;5,BB$.


12/12


Volume 1, Issue 5, Ma !"1#$I%%& !'# #5'

+; C. Fallester, ". Fertalmio, H.Caselles ,G. Sapiro and L.Herdera, b@illing-in )( Loint #nterpolation ofHe&tor @ields and Gra( Je/elsc, #n #EEE ransa&tions on #mage Pro&essing,Hol.! =$>,pp.!!-

,!!.

+? .Chan and L.Shen b9on-e0ture inpainting )( &ur/ature Dri/en Diffusion=CDD>c,L./iusalComm.imagerep.su)mitted!!!.

+$ A. Criminisi, P. Pere6 and K. o(ama, bNegion filling and o)7e&t remo/al )( e0emplar-)asedinpaintingc. #EEE ransa&tions on #mage Pro&essing B,pp. !!-,!!4.

+B L. @lusser, L. Fold(s and F. ito/a, b"oment forms in/ariant to rotation and )lur in ar)itrar(num)er of dimensionsc, #EEE ransa&tions on Pattern Anal(sis and "a&hine #ntelligen&e /ol.:,

pp. 54-4;,!!5.

+! D. Drori , Cohen-

Digital Image Inpainting Using Patch Priority Based Method

Documents