Module-5 DIP

Resiaza 1ion

’ Model o Imaqe Deqradation / Restovation Ps0cess

No ise m od et
Regto ation 4he Presence o Noise onty us fng
Spatial Filteing and
Fvequency Domain

Inverse Fillerinq
Minim um Mean Square ErT0s | wienes] Fiteoin4

Image Rstoration.
Restona 4ion
Image poe - processinq method that

Supress Known deqoadalion g

En hanc ement Restovation

t qives bet{e* visual Temoves eets o

env?0n ment
Tepresentation sens ing
No mode
+ Mathema?cal model o4
Tequired d eqradat ion is need ed

t s an obje «t fve poocess

subject ive procer
(ontaast stetching , histoqra invere iltening wien
equatization ete ave jfMeain q , deno'sîng aoe

Some enhune ment Some e3to at ion

techniques techniquer

ModeloImaqe DLqradaifon Resloration

Degoaded &
nofsy imuqe
Deqradat Ton 9l,4) Restorut ion
Function +
input h(*, 4) degvudrd e3101ed
image image Noise îmage
n(,4)
Deqnada tion Resioration
| Ds. S alee m S Tevavamani
 f(X,y) : osiglnat imuge
h(x,y) deqrad ed junt fon
nlx, ) : add? Live nofse
9(x.4) : degadrd and no isy ima qe
The degrada tfon p0cesR mod eted aß a deqvadation
junction that toqethei with an additive nofse term
opeai es On an input tmuge (6 (3,y) to produce a
dequd ed
the obfe ive 1estovati on is 40 obtain estimate
the oniginal fmage
The degradd îmaqe fn spatial domain is qiven by

* - fn dicates con votution

The equiva lent 7equenty dom ain 7epresentatfon it

4cu,v): Hlu,v) E(u,v)+ N(u,v)

the more we Know a ba ut H and n the
* In qeneval
Close be to

The 7estora lion approach used thsoughout

Testovat ion jflievt
On Various typs o imaqe

’Noise..Modeie
and FTequency propevt iee Nofse
Spaifal
Norse assu med 40 be white notse i. e ouriev
Spectaum o nofise constant

Nofse fs assumed be fndepen dent s patial

domain i.e thee (oVrelatfon between pixet
Value os Tmaye dnd Value o noise compornent
 DIP
S ome imporiant noise20habilitq. Dengity untion.
Noise Which spatia lly depend ent spatia0
tocation is peniodic noise

Fotlowing he most
com monly oc(us ing noise modele

I. 4aussian noise
4aussi an noise models used bequently in pradice
in both
be cause mathena t ical touctubi tity
domain
spatial and bT equ ency
The 4uussiaun 0ndom vasiable z is 4Yv Pn by

(z -Z)*

NI
N
Whee
intensity
mean Vatue o4 Z
1
s tandad deviat fon
P(z)
P Lot P (2) w..t is shoon
Yalues

0.604
he Tange I (Z -r), (~+ r]
and q5 . the Values
the Tange

2.Raylcigh noise,
The PDF o Royleigh noise is qiven by
(2-a)ee-b fo 2 a

fo 2<a

Vunlaoce
The meun and
dentity is given by
and b(4-) /
D Saleemn Tevava mani
 A h is 6quse showe the ptot o
Rayleigh denstty
Raytigh cuve doesn't stavt (1om osiqin
and it is not symmrbial w..t
(Uve

+ Rayteigh densits is use7ul (os

approximat ing SKewed histog dme

3. ilang (q4nMalnoise
The PDF ol Estang noise fs qiven by

P(z) : (b- d!
For Z<0

4he poamnete ane such that a>o b is posttive

whee

integes
! - fndi cat es bactoiat
mean
and vaniance o his denstiy is qfven by
The
6

when Hhe denomina tov is a

P(z)
9amma bunction , The PDF
desribes he qanma
distalbu tfon Plot as

k: a(b-)(b-1) Shown fn the tquse

(6-1)!

(6-1)/a

4 -xponeoLialncise
The P DF exponent lat nofse fs glven by
aeaz
Plz)
$0 z<o
 DIP
whee a >o The mean d nd variance o th is density
6untion
Plz)
This P DF is Specia case o< the
exponential POF with
Ertang
Fíquve shows a plot o thfe
density buntion

5 Uni<o maoise
The PDE o unilom noise fs qiven by
P(2) = b a
othes wise

The mcan this density bunctfon fs qiven by

a+ b (6 - a)

P(z)
This 1tguxe shows the ptot o

b-a uni<osm unilovm density

Impulse (Salt dnd pepPe oise

The PDF fmpu lse noise fs qtven by
Pa
Pl2):

o theu fse

* 4 b a , fn tensiiy b' wile

fmpu ise
appta as a iqht dot on the Pmage
Pa and dak dot
appea ds

N
Th is s bip ctov noise

Dr. 8aleenm Teva ra mani

 Pa 0 Pb 0 unipolas noise

t Pa 0 and onty existe this is calted pe ppex

black dots avt vis ible noise
nofse onty
and only Pa exfte hfe is Called &ult noise

white dot vicible as noise

as only

Peatodic Noise.

dep en dent nofsc

Peiodic nofse
spat?atty
acquis?tion electoical electso chem
During fmage
-ical înteo[esence may (quse type
periodic noise
the
A sto on 4 peoiodic nof se Can be 8een

4e4uency dom ain equi - Spaced dote at

Paoticu las 7adiue ad0und 4he cente o the s pedoum

ES4fmatfon o Noise Pad mete

The pad meteo peoiodfc nofse anl es îmated n
i3e4uen(9 doma in the Se nsoY cpecisf ca ti one ne
Kncwn the noise PPF be KnouUn þoom tt .
Consfdes Hhe matl denoted bs 8

Let Ps ( Zi), ? 0, ,R.. .. L-| denot es the probu bilit

estimates intengít ies o pixelt in s
then mean 6 Vavlance is

gfven
zips( zi)

90y level oj pixete fn sub fmaqee

 DIP
Restoration fn 4he PoesenceoNoise on Lsusing
Spalialiieain4. and Erequency domafn j1Iteving.
when on ty deqrada tion present ?o a fmaqe is
nofse 4hen

g(.9) f(r,y) + n (7,4)

in Fequtncy doma?n

GCu,v) : F(u,v) + N(us V)

Noise Removal meihod

Mean .0 des Statistics Adapt?ve

I. AiHhmetic
I. Medidn ite . Adaptive , tocat
mean 4?!ie
and mîn notse Teduct ion
R. Max
2. Geometuic
mean fHe

3. Hamon fc 3. Midpo tnt iHes R. Adaptive , med ian

Meun bi lier
H. Alpha - 4rimmed
Conia hosmonic meun 6flter
mean ilte3

M2an ieoe.
.Azilh metic mean Tliee
1epresents the coovdina ies n a ectan
Let
gu ior ubimage sze mxn centeved at poin t(x, y).
the on1hmetic me un (omputet avesage
Volue (o1upted Pnage 4(.y) n he debined
by Szy

(9,1) ¬Sey
and
(?ifes &moothf to cat vasia tiont in dn imogt
A meon

nofse is Ted uc ed as nesutt o b lu ing

D Saleem S Tevaramani
Ex (ongide7 sulb fmaqe s(x,y) 0de 3x 3
30 X X

47

30 X

m 3 n:3

3X3
3o +10+ Ro +10+a50+ 25 +R0t R5 +30]

-Geometoic. meun

îmage nestoed using qeometríc mean 6itos fs

gven by expoess?on

Hen ch each 1et0ed pixet ie qiven by the PToduct o< the

pixels in the Subimage , u ised to 4he poweY Im
Ge ometic Mean 67ter uchievPs SMoothinq cOmpundble to
anithm ttic mean 6iltex bu t ?t ends to tose ese

detail io the procese

Rmaqe
Ex Conside un îmage 3X3

|6 X X X

18 X X
x
3

"/3x3
( 5% l6 x2X6X 3x 18 X 1k X3X 15 )

Hamonic.
The hamonic mPon tihoIng oprution t gtuen by
mn

(S,)eSzy 4(c,)
 DIP
hav monic mean 67ter w0Ke wel or salt noise,
nofse . t does well als o with the
but lor pepper
0lhex typee nofse 4?Ke Gauesian nofse

30

250 20

30 X
R0

3 X 3
-
+ 30

|9.91 20

4. Coodaa hasmon.lciler
yietde a esloved fmuqe based on
The contq hus monic i1teY
4he expvessfon 3(s, t ) @tI
(S,t)¬ Sz4

9(s,t)a
(s,)eSy

whee (alled 4he o der o he 61Hes. This 8tes

e timinating the
we! Cuiied sor Te ducing or vtotually
noise
catt - dnd- pe ppe
?4ex eliminutee pepenoige
4he
Cot + ve valu o salt poYge
-Ve vufus 4he ther eliminate
oithnet ic meun

The
(ontra hormonic |Ites educr to
the h avmon ?e
and

16
Q: 0 100

15
+c)'+ ()+ (6 t (3 y+((8y+ 6z' +(3}+ (5)'
4hen (12)°+ (3)
(6)° + (6)° + (22)°+ (6)°+(3)°+ (4)+
+ (15)
00

D Sale m Tevaa mani

 Q 1

I5.12

I5

0rder Statis tic

These non- lineay epatia( jiltese

Median
ilte is the
the beet - Known odes - Statis tic
median 1te
medion g(s,t)?
than Lîntas
7emoVes noise with
bluing
Gmoothing
This 6T}tes ellective bípoBas and unipolar
fnpulse nofse
RO RO 25) 25 30 30 R50

EX
20 X X
30
R50 25
X
| 20 &5 30

Muxaod Mio He
4he 100h pe centile o the
The max

TUn Ke dl numbto!

? (,4) max
(s,t)¬Szy
fs uged (or nding tht briqhtnese point in an
în fma ge has value,

max
 DIP

Min ilter ute7u! or (inding the darkest poini

Ond also it educts S alt noise
?mage
min Te pesenlg 0 th per centile o the unked
The
numbe

min
(s, t) ¬ Sxy

30 » Aranqe all 4the pixel volues în

250
ascending 0rdes Hhen ta Ke the

R5 30 max and min value

x 20 R5 R5
X X x X X

X
30 30
259)Max
X x X R50
X X X X
X

min itter max hi ltes

3. Mid-pointJlLe
The mid point 6ilter cimply com puter the mid point blw
the maximum and minimum Valuee n the

encompass ed by the
min
? (*,9): max
Is,t)¬Sz4 ig(s,t)3 + (s,t)eSz4
fttes Com bines the 0Tder Statieticg and avevage
dictafbuted noise
LwOYK3 best lo 1andom ly
Gauss ion unijovm notse
30 20
X

R60 25 X 130

25 30 X

+[t50+ io]
|30

Saleem Tevuamani
4. Alp hahimnmed

Suppose we d elete the d|2 lowest dnd Hhe d|a hightst

in tensity Valuet 9(s,4) in 1he neiqhhor hood Sxy
epoesent he Te maining mn - d pixe0s
Let 9(z, t)
the síHes bonmed by avesaqing thee nemainfng pixets
is (allod alpha (nimmed jiHe3

mn -d ($,t)¬S1y
mn -I
The value o< d
d 0 4hen al pha trímmed $iMer An4hmetic
d mn-I Alpha ta fmmed jHer = median iltes

FoT d & we Temo ve d. - . 1 min value (i,e I0 in

example ) and 1 man Value (i,e R60 in the below ex)
below

X
3 0
250 26 R3
X
Ro 30

I | 3 0 + & 0+t0+ 26 + 20+ 25 t ?o|

This ites TeMOVes the combfnation salt peppe

and qau esfon noise

Adap4iv
Adaptivt to capable o< supevior penjos mance
because fis behaviou adopie lo the changt în
Chura1eoistis ol fmago area bttnq ilteved But his
inotases the (omptexfty
fn
wo 1ypes Adapiive tens
Adapiive , toal noise 1educt fon
A. Ad aptive Medlan
 Adap tivlocolnoiseuduetton iHx
1his opeoate On
local 1tqion S4 . The 3sponse
at any point (,y) on hich the atg'on

cente ed is to bt based on $ous quoniiHiee

Value o nofsy imaqt at (1,4)
ihe Vaniance nofse co7upt în g f ,y) o
9(,4)
3. m he local mean 1he o pfxet in Szy
the local voviance ol the pf x ele in

ollows
Ihtbehariousot
should eun ef mply he
zev0, the ite
value o 9(,y). 1h is is the iivia|, 2e0 - noice case in
which q(,4) is equal to f(,4)
4Ahe lacal ariance is hiqh elative o Lai the <ilis
Should etun Value close

yhL equal theilter Telun the

axíthonelic 0the ixel in Sxy

? (3.4)

k. Adaptive MediaA iltes

The Adaptive mediun hilte Can handle impu tte
(Ompare to n0 mal medfan te
noise

The mafn objetivt o| adaplive medlan 4Res ie

To Temove and pebpe (impuite ) nof se
Smoothen noise 0ther than tm puie noise

1edue
disior tfon o} thinníng and thtckenYn g
To

Variablee used fn he atqovtihm

Sy = Rectonqu tas wind ow
D Saleen Tevanumani
Z min : minimum qay levet Vulue in
q7ay level Vulue in
Z max maxinum
tevel value în
S4
Z med median 4304
at (x,y)
Zy
m axiwmum allowed
Smax
Adap tive 6ltes wo ke in two aenoted
median ctages
Stuqe A and stage B as (oows

Slaqe A AJ = Zmed - Zmin

AR = Zmed- 2 Max

AR<0 , 90 to staqe B
AND

increaSe window síze

windouw sige s Smax epeat Stage A
output Zmed

Stugt 8, 81 Zy Z min
B Zy - Z max

AND 82< 0 , out put Zz4

E Lse oul put Z med

Domafn
Peledie Nnis Reducian h EIlquency
Jilieníng.
Periodic noíse Can be analyzed and |Heved quite
eljecdively using }equency dom ain 1tchniqutt
is to use elective iHe to isotale
the approach
4he nofse .
the 4hte typts o selective iltht bundveject
band pase und nolch i1t

periodic noise 0M an imaqe in volve

Removi ng (vom the
vtmoving paylitular Yang' o quencee
imag
 The ans|e <unction o ideal band (t
D(u,v)< Do -
H(u,v ) : D- s Dlu,v) Do +

D(u,v) Dot

whee D(u,v) is Ihe aistance the origin and tt

the above equation
qfvtn b4
widlh 4he bund
4he

Do is
adial cent
H(u,v)
V H(u,)

Do +w olu,r)

Tesponse ideal band Tjtet |iHey

Tht ivansje unction o buiitt ulor th hund eect lite
0der is 9fven by
H(u,v ) :
Dlu,v) w
D² (u,v) -D.
3. |for GoSsíu. hand 1Ëecd truns| juntion ir given
by H,lu,v) : |- eaD(u,) -D
D(uv )w

2. and Passilter
The Bund past |te p o m the OPposite opevalfon
o a band e íect
The Irann unctlon Hapu,v) o< a band paes |7les it
obiained
(ovvespond tng bund ejett iltes wtlh
honsy junttion HR(u,v)
8uleem Tevamani
 I- HeR (u,v )

tans je undior o bundpuse 1itte1

HBp -
unt|e untlon oj band ve( ect iltey

Ideal band.pusiRe_s qiven by

D(u,v) s D
H (u,v) Do- < olu,v)< Do t

whee D(u,v) the distance (7om onigin

= Band widh
Do = ad ial (ente

2 Butiewoth band pass iHez o odes' n' is glven by

H(u,v):I 2n

D(u,v ) -D,

3 Simi tus ly Gausstun bund pass

(1He3 is qfven by
Hlu,v) : - H(u,v) G8e
D*(u,v) - Do
D(u,v)w
H(u,)
H Cu,v)
H (4w

olu)
Da tw

Fitqutny ideal bund

 DIP
3. Nochilles
A |i ies 1e[ecie ( o paste)
notch j1es requtnt fee fn
Ptdelintd nelgh bov hoods about a (ente j2quenc4

HNp( u,V) is the dns|e* {unction ol notch puss LTHes

whee
fs the transe unction o notch Teject
HNR (u,v)
|tes
Nadch aict
L* The Tiansle (unction o Ldea! notch I;ct yiler with
(uo, Vo) and (- uo, Vo) is
Diluv) < Do Da (u,v) < Do
Hlu,v):
oHherwige
Cuo,Vo) & (-uo,-Vo)
D.(u,) is +he disiance 7 om
wh ere Dilu,v) and
Tadiu
Do

D1 (u- -u)'+ (v- 4-vo)|

butler w07 th notch Tejet iter ol ovdes

Transle unction o
n is q?ven by

H (u,v) 2 D

G4 ussian aatch eect Ilte is qiven by

H(u,v): I- | 9 )Da (u,v)
Do

4 uo = Vo s 0 all 4hee itteo become high pase 6itet

(-Mo,-Vo)

Frequency Teshonre o| 1deal notch vefect (?#es

D Saleem Tevasamani
 Notch Pass ilHess
Tansjer unct ion oj notch pass Thes can be obta ined T0m
notch eject hiltens
Hnplu,v) = |- Hnglu,v)
DI (u, v) < Do o7 Da (u,v) sDo
i.e H(u,v) :
otheowise

(-us,-Vo) (us, Vo)

Frequtny neponee ol id eal notch pus iHes

Buttew0th hotch puss 6iHes
Trons4es untion o

Do

Gau gsian notch puss iRes is gtven by

H (u,v) : |- H( MV)NPF

ilHess be come low pass iltes i| Uo = Vo 0

Notch pass

Notch Fileoinq
7he gtep fn optimum notch itter is to extacd the
Princip!e (0quency compone nt o the inieetnce paifeon
|0m nofsy imagt
Th ?s done by ploclnq a hotch pace iHey #Cu,v)
at Hhe to cat ion o, each nolse cpke
Founfer t ont le m o intelesence noise puttevn it qiven by
Nu,v) : H (u,v) G(u,v)0
whee
Glu,v): Founter tdns lovm o co11upted image
H(u,v) optiMum notch lliP ane (unct ion
 DIP
(an be wi11en
Ln spatiat doma fn
n , y) G(u, v)]
F' H(u,v) G
no9): [ H (4.v) J. 9(*,9)
Cor upted Imaqe g(9) i: a ssumed
to jormed by
be

addiion unco upted fmage f (.9) and noise n(4)

4he noise 1e3m n(9) were Known completely it (an

but i the nofse is an

0Ppvox f mallon 4hen wefghted portfon o n(yy) is

Substacted 67om f(zy)

ihe estimate
w(X, 4) is to be de temined

+ Cong i des ne fghbo hood o< eize ( 2at ) by c Rb+1) about a

point (*9). The local Vaiance o fCzy) at co0vd inatc
be est fw a ted 4 he
(*.4)
t: -b
(*+s, yt t) - f(4, 9) -6
(kat)(Rb+1)
whese is the aveage value o

(Ra+i) (26+1) S:-a

substilutBng
(Rat) (Rh ti) Se -a

Acsum in g wlay) ma fng Constant ntigh boshood

w(* +S yti) : w(7, y) )
dnd - 6< t sb this ascum ptfon
4he exprtssPon
w (9) n (a y)
D Suleem Tevu amani
whee wl,y) aveDage Value w(y)
be (omes

r'l,y):
(Ra+)eb+) E { 915,ytt) - wCzy)n (1+5,yri)|
S: -a

minimige 0
dw ( y)
Fo w (xy)

w(Y)

(illering
TestoTa tion is direct inverse
The sinplest approach to
whee compute an estimate

transo m o 4he orig'nal fma ge by aivtd ?ng

the
the 4he deqruded Tmaqe
deqradat f on 6undfon
G (u,v )
H (u,v)

whe F(u,v) : esto ed fmage

Founier tans7om o< oríginal imaqe F(u,)
H (u,v) : degadat fon Junction
w. k.T G Cu,v) = H (u,v) F u,v) + N (u,v)
E (u,v) + N(u,v)
H (u,v)

One approach to avound the Sm al< -va lue

problem fs to 4he ITequen ( fes to value

nea Ahe ofgfn

the
thoughdegu dailon juntfon H(u, v) is Known
4he unga ded fima ge hot bl obta fn ed
completety beause
quantity squared . 1h is lsult as he wieneY

H(ua)i deqna dati on junction.

H* (u,v) : (Omplex (onju gaie o H(u,v)
H*(u,v) H(u,v)
1he n oise
S Cu,v)ELus)powex undeqraded image
wiene lorm as in eq disicut to
esiim ate Sn (u,) dnd Srlu,v)
To eolve do appox imatiOn
Sn (u,v)
Ss (u,v)
Then becom es
IH Cu,v))*
G(u,v)
K chosen experimen tally
In O des to |ind Sn (u,v) and S (u,v) the
used
jotlowing
1 Siqnal Ho noise atio (SNR)
N-! N-I

SlN( u,v) |*
lmaqes with low noise has lower S NR

Mean 4uuae e330Y(MSE)

M-I
MSE =
MN
we COn
de7ine he ergnat to nofse tatio fn the
spatial dom afn al

Dy Saleem S Teva amani

. 1he nofse tompon ent m
und om jundion wheoe

lounies hons [ov m N (u,v) not Known

In pvactice H(uV) (an (ont ain nUmeYOU zeos

’Min in umMean square ( wiene)EMeI inq

wiene ilte estores 4he im a qe fn the presence oj

blus weLl ag nofse
to fnd an estimae the
The objective
that mean
unco upted fmagt f such he

blw 4hem mfnfmige d This is given

by
Where the expected Value

undeqrg d ed im a qo Tcovupted image1

aumed that the noise and the imaqt
is mean and
other bas
hat
un co de la ted i estimate oinea
în the
înten si t4 levels
that +he
tevets fn deqraded image
fundion
(ondiliong 4he m înimum o the
Based On these domain
Iuntion is qven in the 74u Pn(9
by the ex pvession
H* (u, v) S (u,v)
(u,v) :
Sf (u,v) | H(u,v)j2+
Snlu,v) |ctuw)
H (u,v)

G(uav)
Ss (u,v)

The act ihat 4he product o a complex quan ity with

is equat 1o 4he maqnftude o< comp!x
(on fu qate
 M-I

SNR :

M-I N1

IH (u,v) )
H (u,v) Glu,v)
|H (u,v)j² +K
(onst ont that a dd ed to ail
whese
tem o IH(u,v)|*.

. For a qiven image detevmfne Z min , Z max Zmed Zzy

qnd Sma X

3 6

13

Soul Z mfn : 1

Z med I 3 6 (8) 4 I| 13 5

[ nteas ity valt at cooTd ina es

S max : 3 X3 | max a tlowed size

D Saleem Tevu amani