notes7 image enhancement i - university of arizonadial/ece533/notes7.pdfece/opti533 digital image...

28
ECE/OPTI533 Digital Image Processing class notes 137 Dr. Robert A. Schowengerdt 2003 IMAGE ENHANCEMENT I (RADIOMETRIC) MOTIVATION • Recorded images often exhibit problems such as: • too dark • too light • not enough contrast • Image enhancement aims to improve visual quality “Cosmetic” processing • Usually empirical techniques, with ad hoc parameters (“whatever works”) • Distinct from “restoration,” which is processing designed to recover true object using blur and noise reduction on a degraded image

Upload: phamkien

Post on 23-May-2018

223 views

Category:

Documents


0 download

TRANSCRIPT

EC

E/O

PT

I533 Digital Im

age Processing class notes 137 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

MO

TIVA

TION

•R

ecord

ed

ima

ges o

ften

exh

ibit p

rob

lem

s s

uch

as:

• to

o d

ark

• to

o lig

ht

• n

ot e

nou

gh

con

trast

•Im

ag

e e

nh

an

cem

en

t aim

s to

imp

rove v

isu

al q

ua

lity

“Cosm

etic

” p

rocessin

g

•U

su

ally

em

piric

al te

ch

niq

ues, w

ith a

d h

oc p

ara

mete

rs (“

wh

ate

ver

work

s”)

•D

istin

ct fro

m “

resto

ratio

n,”

wh

ich

is p

rocessin

g d

esig

ned

to

recover tru

e o

bje

ct u

sin

g b

lur a

nd

nois

e re

du

ctio

n o

n a

deg

rad

ed

im

ag

e

EC

E/O

PT

I533 Digital Im

age Processing class notes 138 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

IMA

GE D

ISPLA

Y

•In

pu

t qu

an

tized

ima

ge p

ixel v

alu

es (in

teg

ers

): Dig

ital N

um

ber

(DN

)

•O

utp

ut q

ua

ntiz

ed

ima

ge p

ixel v

alu

es (in

teg

ers

): Gre

y L

evel (G

L)

GLs a

re in

dis

pla

y s

pa

ce, ty

pic

ally

[0..2

55

] in e

ach

colo

r (red

, gre

en, b

lue)

DN

1LU

TGL

RD

/A

RGB

DN

2LU

TGL

GD

/A

DN

3LU

TGL

BD

/A

RGB

colo

r im

ag

e

EC

E/O

PT

I533 Digital Im

age Processing class notes 139 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

•M

ap

pin

g fro

m D

Ns to

GLs m

ay b

e d

on

e w

ith d

iscre

te h

ard

wa

re o

r softw

are

Look-U

p Ta

ble

s (L

UTs

)

•Con

trast e

nh

an

cem

en

t tech

niq

ues a

re d

esig

ned

to fin

d a

LU

T tha

t yie

lds “

op

tima

l,” o

r at le

ast “

good

,” d

isp

layed

vis

ua

l qu

ality

Resu

lt is a

rad

iom

etric

en

ha

ncem

en

t of th

e d

isp

layed

ima

ge, i.e

fea

ture

s in

the

ima

ge b

ecom

e m

ore

ea

sily

seen

GL

LU

TD

N(

)=

EC

E/O

PT

I533 Digital Im

age Processing class notes 140 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

IMA

GE S

TATIS

TICS

•Glo

ba

l ima

ge D

N s

tatis

tics u

sefu

l fo

r desig

nin

g a

n im

ag

e-s

pecific

LU

T fo

r en

tire im

ag

e

•Loca

l ima

ge D

N s

tatis

tics u

sefu

l for

“a

da

ptiv

e” c

on

trast e

nh

an

cem

en

t a

lgorith

m, w

ith v

ary

ing

LU

T acro

ss

ima

ge

•N

eig

hb

orh

ood D

N s

tatis

tics u

sefu

l fo

r nois

e p

rocessin

g a

nd

sp

atia

l filte

ring

en

ha

ncem

en

t, not c

on

trast

en

ha

ncem

en

t“g

lob

al”

“lo

ca

l”

“n

eig

hb

orh

ood

EC

E/O

PT

I533 Digital Im

age Processing class notes 141 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

Ima

ge H

isto

gra

m

•N

um

ber o

f pix

els

with

the s

pecific

D

N, ta

bu

late

d fo

r all D

Ns

•D

ivid

e b

y th

e to

tal n

um

ber o

f pix

els

in

the im

ag

e N

to n

orm

aliz

e

An

alo

gou

s to

the c

on

tinu

ou

s P

rob

ab

ility

Den

sity

Fu

nctio

n (P

DF) o

f sta

tistic

s

•Con

tain

s n

o in

form

atio

n a

bou

t the

sp

atia

l dis

tribu

tion

of p

ixels

histD

Npixel count

DN

()

=

no

rmh

istDN

countD

N(

)N⁄

PD

FD

N(

)≈

=

Exa

mp

le n

orm

aliz

ed

ima

ge

his

tog

ram

com

pa

red

to th

e

eq

uiv

ale

nt G

au

ssia

n

dis

tribu

tion

0

0.05

0.1

0.15

020

4060

80100

Gaussian

image

fraction of total pixels

DN

µµ

+ σµ

− σ

EC

E/O

PT

I533 Digital Im

age Processing class notes 142 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

Cu

mu

lativ

e H

isto

gra

m

•N

um

ber o

f pix

els

with

a D

N le

ss

tha

n o

r eq

ua

l to th

e s

pecifie

d D

N

•D

ivid

e b

y th

e to

tal n

um

ber o

f pix

els

in

the im

ag

e N

to n

orm

aliz

e

An

alo

gou

s to

the C

um

ula

tive D

istrib

utio

n

Fu

nctio

n (C

DF) o

f sta

tistic

s

•M

on

oto

nic

fun

ctio

n o

f DN

chistD

Nh

istDN

DN

DN

min

= DN

∑=

no

rmch

istDN

chistD

NN⁄

CD

FD

N(

)≈

=

Exa

mp

le n

orm

aliz

ed

ima

ge

cu

mu

lativ

e h

isto

gra

m

com

pa

red

to th

e e

qu

iva

len

t Ga

ussia

n C

DF

0

0.2

0.4

0.6

0.8 1

020

4060

80100

Gaussian

image

fraction of total pixels

DN

EC

E/O

PT

I533 Digital Im

age Processing class notes 143 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

RA

DIO

METR

IC E

NH

AN

CEM

EN

T

•Poin

t pro

cessin

g

every

pix

el u

nd

erg

oes th

e s

am

e

tran

sfo

rma

tion

pa

ram

ete

rs o

f the tra

nsfo

rma

tion

are

:

• fix

ed

, from

glo

ba

l sta

tistic

s (e

ntire

ima

ge) o

r

• a

da

ptiv

e, fro

m lo

ca

l sta

tistic

s

•D

istin

ct fro

m n

eig

hb

orh

ood

p

rocessin

g (c

on

volu

tion

) or g

lob

al

tran

sfo

rms (e

.g. F

ou

rier)

•O

nly

imp

roves v

isu

al c

on

ten

t; does

not a

dd

info

rma

tion

to th

e d

ata

EC

E/O

PT

I533 Digital Im

age Processing class notes 144 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

Min

-Ma

x M

ap

pin

g G

L255

DN

ma

xD

Nm

in–

--------------------------------------D

ND

Nm

in–

()

=

050

100

150

200

250

300

0

2000

4000

6000

8000

10000

12000

14000

DN

050

100

150

200

250

0

2000

4000

6000

8000

10000

12000

14000

GL

DN

ma

xD

Nm

in

DN

GL

255

0D

Nm

ax

DN

min

EC

E/O

PT

I533 Digital Im

age Processing class notes 145 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

His

tog

ram

Mod

ifica

tion

•n

on

linea

r “stre

tch

ing

” o

f the d

ata

to im

pro

ve c

on

trast

ba

sed

on

glo

ba

l or re

gio

na

l ima

ge s

tatis

tics

EC

E/O

PT

I533 Digital Im

age Processing class notes 146 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

His

tog

ram

Eq

ua

liza

tion

•M

od

ify h

isto

gra

m to

ach

ieve u

nifo

rm d

istrib

utio

n o

f GL

•Look a

t con

tinu

ou

s d

istrib

utio

ns fo

r “p

roof”

Ma

pp

ing

fun

ctio

n fro

m in

pu

t to o

utp

ut:

Ou

tpu

t PD

F re

late

d to

inp

ut P

DF a

s:

desire

ou

tpu

t PD

F th

at is

flat, i.e

. un

iform

dis

tribu

tion

con

sid

er a

tran

sfo

rma

tion

M th

at is

the C

DF o

f x: (w

e k

now

the a

nsw

er!)

yM

x()=

PD

Fy()

PD

Fx() d

xd

y------

xM

1–y()

==

yM

x()P

DF

α()

αd∞– x∫

==

EC

E/O

PT

I533 Digital Im

age Processing class notes 147 D

r. Robert A

. Schowengerdt 2003

IM

AGE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

then

su

bstitu

ting

into

ab

ove e

qu

atio

n fo

r PD

F(y

):

•D

iscre

te c

ase is

sim

ply

,

dy

dx

------P

DF

x()=

PD

Fy()

PD

Fx()

1P

DF

x()--------------------

⋅x

M1–

y()=

=

1[]

xM

1–y()

==

1=

GL

int

255ch

istD

N(

)[

]=

DN

GL 2550

DN

ma

xD

Nm

in

EC

E/O

PT

I533 Digital Im

age Processing class notes 148 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

ou

tpu

t his

tog

ram

050

100

150

200

250

0

2000

4000

6000

8000

10000

12000

14000

GL

EC

E/O

PT

I533 Digital Im

age Processing class notes 149 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

His

tog

ram

Ma

tch

ing

•M

od

ify h

isto

gra

m to

ma

tch

“re

fere

nce” h

isto

gra

m, e

.g. G

au

ssia

n

•Ca

n a

lso b

e u

sed

to m

atc

h h

isto

gra

m o

f on

e im

ag

e to

tha

t of

an

oth

er, e

.g m

ultid

ate

ima

gery

of s

am

e s

cen

e

GL

no

rmch

istref 1–n

orm

chist

DN

()

[]

=

DN

0 1

normchist

DN

ref

0 1

normchistref

EC

E/O

PT

I533 Digital Im

age Processing class notes 150 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

ou

tpu

t his

tog

ram

ma

tch

ed

to G

au

ssia

n

•Pro

du

ces “

softe

r” c

on

trast th

an

his

tog

ram

eq

ua

liza

tion

050

100

150

200

250

0

2000

4000

6000

8000

10000

12000

14000

GL

EC

E/O

PT

I533 Digital Im

age Processing class notes 151 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

AD

APTIV

E C

ON

TRA

ST S

TRETC

H

•A

da

pts

to “

loca

l” c

on

trast

diffe

ren

ces

•R

esu

lting

con

trast is

more

un

iform

a

cro

ss e

ntire

ima

ge

EC

E/O

PT

I533 Digital Im

age Processing class notes 152 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

Loca

l Ra

ng

e M

od

ifica

tion

(LR

M)

•Pa

rtition

ima

ge in

to a

dja

cen

t b

locks, X

x Y

in s

ize

•Fin

d th

e m

inim

um

DN

L a

nd

m

axim

um

DN

H in

ea

ch

blo

ck

MIN

1 ,M

AX

1

MIN

2 ,M

AX

2

MIN

3 ,M

AX

3

MIN

4 ,M

AX

4

MIN

8 ,M

AX

8

MIN

12 ,M

AX

12

MIN

5 ,M

AX

5

MIN

6 ,M

AX

6

MIN

7 ,M

AX

7

MIN

11 ,M

AX

11

MIN

10 ,M

AX

10

MIN

9 ,M

AX

9

MIN

13 ,M

AX

13

MIN

14 ,M

AX

14

MIN

15 ,M

AX

15

MIN

16 ,M

AX

16

X

YL

A,H

AL

B ,HB

LC ,H

C

LG

,HG

LE ,H

EL

D,H

D

LH

,HH

LF ,H

F

LI ,H

I

y

x

EC

E/O

PT

I533 Digital Im

age Processing class notes 153 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

•Fin

d th

e n

eig

hb

orin

g m

inim

um

an

d m

axim

um

L a

nd

H v

alu

es a

t th

e in

ters

ectin

g n

od

es, e

.g.

•In

the c

orn

ers

an

d a

lon

g th

e e

dg

es, e

.g.

,

MIN

6m

inim

um

LA

LB

LD

LE

,,

,(

)=

MA

X6

ma

ximu

mH

AH

BH

DH

E,

,,

()

=

MIN

1L

A=

MA

X1

HA

=

MIN

5m

inim

um

LA

LD

,(

)=

MA

X5

ma

ximu

mH

AH

D,

()

=

EC

E/O

PT

I533 Digital Im

age Processing class notes 154 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

•Lin

ea

rly in

terp

ola

te th

e e

xp

ecte

d o

utp

ut G

L m

inim

um

an

d

ma

xim

um

at e

ach

pix

el fro

m th

ese v

alu

es

•Lin

ea

r inte

rpola

tion

insu

res th

at w

e’re

alw

ays w

ithin

a s

pecifie

d

ou

tpu

t ran

ge, e

.g. [0

..25

5]

GL

min

xX ----M

IN7

⋅X

x–X

------------

M

IN6

⋅+

Yy

–Y------------

⋅=

+

xX ----M

IN11

⋅X

x–X

------------

M

IN10

⋅+

yY ---⋅

GL

ma

xxX ----

MA

X7

⋅X

x–X

------------

M

AX

6⋅

+Y

y–Y

------------

=

+

xX ----M

AX

11⋅

Xx

–X------------

MA

X10

⋅+

yY ---⋅

MIN

GL

min

MA

X≤

MIN

GL

ma

xM

AX

≤≤

EC

E/O

PT

I533 Digital Im

age Processing class notes 155 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

•Fin

ally

, ca

lcu

late

a lin

ea

r stre

tch

at e

ach

pix

el u

sin

g th

e

estim

ate

d m

inim

um

an

d m

axim

um

ran

ge v

alu

es

•Sen

sitiv

e to

blo

ck s

ize

too la

rge p

rod

uces little

ad

ap

tivity

to lo

ca

l con

trast

too s

ma

ll pro

du

ces “

ha

lo” a

rtifact n

ea

r hig

h c

on

trast b

ou

nd

arie

s

find

“b

est”

blo

ck s

ize b

y tria

l-an

d-e

rror

•R

efe

ren

ce:

J. D. F

ah

nesto

ck a

nd

R. A

. Sch

ow

en

gerd

t, "Sp

atia

lly - v

aria

nt C

on

trast E

nh

an

cem

en

t u

sin

g L

oca

l Ra

ng

e M

od

ifica

tion

,” O

pt. E

ng

., 22

(3), p

. 37

8 - 3

81

, Ma

y - Ju

ne 1

98

3.

GL

255G

Lm

ax

GL

min

–------------------------------------

DN

GL

min

–(

)=

EC

E/O

PT

I533 Digital Im

age Processing class notes 156 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

IMA

GE E

XA

MPLES

orig

ina

l

EC

E/O

PT

I533 Digital Im

age Processing class notes 157 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

min

-ma

x s

tretc

hed

EC

E/O

PT

I533 Digital Im

age Processing class notes 158 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

his

tog

ram

eq

ua

lized

EC

E/O

PT

I533 Digital Im

age Processing class notes 159 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

Ga

ussia

n s

tretc

hed

EC

E/O

PT

I533 Digital Im

age Processing class notes 160 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

LR

M s

tretc

hed

(blo

ck s

ize =

10

0)

EC

E/O

PT

I533 Digital Im

age Processing class notes 161 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

LR

M s

tretc

hed

(blo

ck s

ize =

50

)

EC

E/O

PT

I533 Digital Im

age Processing class notes 162 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

CO

LO

R IM

AGE C

ON

TRA

ST

EN

HA

NCEM

EN

T

•Ca

n a

pp

ly m

on

och

rom

e te

ch

niq

ues

to e

ach

colo

r com

pon

en

t (“b

an

d”)

•Som

etim

es le

ad

s to

un

desira

ble

colo

r sh

ifts a

nd

poor c

olo

r ba

lan

ce

•M

ore

late

r on

colo

r ima

ge

en

ha

ncem

en

t . . .

EC

E/O

PT

I533 Digital Im

age Processing class notes 163 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

exa

mp

le o

rigin

al

EC

E/O

PT

I533 Digital Im

age Processing class notes 164 D

r. Robert A

. Schowengerdt 2003

IMA

GE E

NH

AN

CEM

EN

T I (RA

DIO

METR

IC)

his

tog

ram

eq

ua

lized

in e

ach

ba

nd