bright_inversion

8/9/2019 bright_inversion

1/23

-186-

Chapter 10 Image Processing - Conversion

10.1 Image Enhancement and Feature Extraction

Image enhancement can be defined as conversion of the image quality to a better and

more understandable level for feature extraction or image interpretation, while radiometriccorrection is to reconstruct the physically calibrated value from the observed data.

On the other hand, feature extraction can be defined as the operation to quantify the

image quality through various parameters or functions, which are applied to the original image.

hese processes can be considered as conversion of the image data. !mage enhancement is

applied mainly for image interpretation in the fom1 of an image output, while feature extraction

is normally used for automated classification or analysis in a quantitative form "see #igure

1$.1.1%.

a. !mage &nhancement

ypical image enhancement techniques include gray scale conversion,

conversion, color composition, color conversion et!een "#$

etc., which are usually applied to the image output for image interpretation.

histogram

and %&I,

b. #eature &xtraction

#eatures involved in an image are classified as follows.

"1% 'pectral features

special color or tone, gradient, spectral parameter etc."(% )eometric features

edge, linearment, shape, si*e, etc.

"+% extural features

pattern, spatial frequency, homogeneity, etc.

#igure 1$.1.( shows three examples of spectral, geometric and textural feature extraction.


2/23

-(

!mage

)ray scale conversion

"istogram conversion%

enhancement olor composition

'! conversion

!mage onversion'pectral feature

rincipal components etc.

#eatureextraction

)eometric feature

&dge, /inearment etc.

extural feature

'patial frequency etc.

Figure 10.1.1 Image conversion

a% 'pectral feature b% )eometric feature c% extural feature

Figure 10.1.2 Samples of feature extraction


3/23

10.' #ray &cale Conversion

#ray scale conversion is one of the simplest image enhancement techniques. )ray scaleconversion can be performed using the following function.

y 0 f (x)where x original input data

y converted output data

!n this section, the following five typical types are inrt2oduced, though there are many more

functions that could be used. " see #igure 1$.(.1%

a. /inear conversion

y = ax + b a gain , b offsetcontrast stretch is one of linear conversion as follows.

'tatistical procedures can be also applied in two ways as follows.

(1) onversion of average and standard deviation

- 'y-(x-

,3m

where xm average of input image

'x standard deviation of input image

ym average of output image

'y standard deviation of output image

"(% 4egression

!n such cases as multi-date images for producing a mosaic or radiometric ad5ustment, a

selected image can be related to other images using regression technique.

/ine noise due to different detectors, for example /andsat '', can be eliminatedby using

the regression techniquebetween different detectors.

#igure 1$,(.( shows various examples of gray scale conversion.

b. #old conversion

ultiple linear curves are applied in order to enhance only a part of the gray scale.

c. 'aw conversion

7here a discontinuous gray scale, occurs, drastic contrast stretch can be made.

d. ontinuous function

#unction such as exponential, logarithm, polynomials etc. may be applied.

e. /ocal gray scale conversion

!nstead of the conversion being applied to the whole scene by a single formula, parameters

are changed with respect to small local areas.

3 m% 9


4/23

y

Xmin Xmax x X

(a) linear (b) fold

y y

X

(c) saw (d) continuous

Fig 10.2.1 Typical Density Conversion

a% original !mage b% modified conversion

Fig 10.2.2 Examples of gray scale conversion


5/23

10.( %istogram Conversion

%istogram conversion is the conversion of the histogram of original image to an other

histogram. istogram conversion can be said to be a type of gray scale conversion.

here are two typical histogram conversion techniques.

a. %istogram e)uali*ation

istogram equali*ation is to convert the histogram of an original image to equali*ed

histogram as shown in #igure 1$.+.1. :s a first step, an accumulated histogram should be

made. hen the accumulated histogram should be divided into a number of equal regions.

hirdly , the corresponding gray scale in each region should be assigned to a converted gray

scale.

he effect of histogram equali*ation is that parts of the image with more frequency variation

will be more enhanced, while parts of an image with less frequency will be neglected.

#igure 1$.+.( shows a comparison between the original image and the converted image,

after histogram equali*ation.

b. %istogram normali*ation

)enerally a normal distribution of the density in an image would create an image that is

natural for a human observation. !n this sense the histogram of the original image may be

sometimes converted to the normali*ed histogram as shown in #igure 1$.+.+. owever in

this conversion, pixels with same gray scale should be reallocated to other pixels with a

different gray scales, in order to fomlt a normali*ed histogram.

herefore such a gray scale conversion is not a 11 conversion and thus enables no reverse

conversion. istogram normali*ation may be applied, for example, to an unfocused image

of a planet with a low dynamic range, though it is not be very much popular for ordinary

remote sensing data.


6/23

ff

(a) histogramX

Xmodified histogram

y y

(b) cumulative histogram X linear cumulative histogram X

by modification

(c) original image modified image

Fig.1.!.1 histogram e"uali#ation

f

histogram Fig.1.!.$ X normali#ed histogram X

normali#ed histogram


7/23

10.+ Color isplay of Image ata

olor display of remote sensing data is of importance for effective visual interpretation.

here are two color display methods; color composite, to generate color with multi-band data

and pseudo-color display, to assign different colors to the gray scale of a single image.

a. Color Composite: color image can be generated by composing three selected multi-band images with the use

of three primary colors. , ) and 4 are assigned to the same spectral regions of blue, green and

red as shown in #igure 1$.=.(, almost the same color as the natural scale, can be

reproduced, and is called a natural color composite.

owever in remote sensing multi-band images are not always divided in to the same spectral

regions as the three primary color filters. !n addition invisible regions, such as infrared, are

often used, which are required to be displayed in color. :s a color composite with an

infrared band is no longer natural color, it is called a false color composite.

!n particularly the color composite with the assignment of blue to the green band, green to

the red band and red to the near infrared band is very popular, and is called an infrared

color composite, which is the same as found in color infrared film "see #igure 1$.=.(%.

!n the case of digital data, three values corresponding to4, ) and > will ma?e various color

combinations, as listed in able 1$.=..1.

b. Pseudo Color isplay


8/23

> >lue

) )reen

4 4ed

yan

agenta

3 3ellow

"a% additive color composite "b% subtractive color composite

Figure 10.4.1 et!o"s of color composite

7 7hite

>/ >lac?

Ta#le 10.4.1 Samples of color compositefrom "igital "ata

'pectral reflectance of vegetation

7ave length

% ) & !4 Original >and

:ssignment @ > ). 4of three ------

primary > )colors --

Aatural color composite

4 !nfrared color composite

Figure 10.4.2 Example of color composite

3ellow

>lue yan )reen )reen 3ellow Orange 4ed

$$$$$$$$$$$%&r.%$&$&$&$&$&$&$

,.I '..'.'.'(' '.... ..

agenta

.....

.........%

-'-'-' *

I '..

.i ...'.. ..

olor tone

---R

Figure 10.4.' Example of pseu"o color "isplay

> ) & olor

$

127

255

$

127

255

$

127

255

>lac?

)ray

7hite

255

$

$

$

255

$

$

$

255

>lue

)reen

4ed

255

255

$

255

$

255

$

255

255

yan

agenta

3ellow


9/23

10. Color "epresentation - Color ixing &ystem

/ight is perceived as color by the human eye, termed color stimulus, and corresponds to

the visible region of electro-magnetic spectrum, with a specific spectral curve of radiance as

shown in #igure 1$.B.1.

owever as the physical value such as a spectral curve, is not convenient for representing

color in daily life, a psychological representation or sensitivity expression are morepractical.

olor representation can be classified into two types; a color mixing system using a

quantitative and physical approach, and a color appearance system using a qualitative

approach, by color code or color sample.

he color mixing system can generate any color by mixing of the three primary colors. he

"#$ color system specified by !&, uses three primary color stimuli; blue of =+B.8 nm">%,

green of B=6.1 nm")% and red of C$$.$ nm"4% by which all spectral colors ranging from +8$

nm to C8$ nm, can be generated with the mixing combinations "termed a color matchingfunction or spectral tristimulus values/ as shown in #igure 1$.B.(.

:s a part of the three spectral stimuli r ":%,g ":% and b":% includes a negative region, a

coordinate transformation is applied to generate the virtual three spectral stimuli r":.%,g":.% and

x":%,3":.% and #( : % with positive values, as shown in #igure 1$.B.+. his is called the ()*

color system. he three stimuli, , 3 and D, can be computed as follows.C8$

x":%/":%p":%d:

8$

C8$

=, y":%/":%p":%d:

8$

780

=, *":%/":%p":%d:

8$

where E constant

/":% spectral irradiance of standard illumination

p ":% spectral reflectance of sample

richromatic coordinates "x,y% can be computed as follows.

X X y y939D F 939D

he value of 3 corresponds to brightness while the coordinates of "x,y% represent hue and

saturation "or chrome% that is drawn as a trichromatic chart as shown in #igure 1$.B.=.

he fringe of the bell shape corresponds to the spectral color with high chrome, while the

inside corresponds with a low chrome.

1X=,1

1


10/23

!.' 1-9-1- -9----t----t--5

$

$.(

;.

.&l $.11--1--9---9-11--9-G9----G--t---1

e:::1Hf5l of-=- ...I -9......-...,;-i

E

-o.l . .L -- soo ,oo

7ave length "nm%

Figure 10.+.1 Color matc!ing functionof

CIE1,'1-/

soo

7ave length "nm%

Figure 10.+.2 Color matc!ing function of

CIE1,'1()*

B($nm

y

>luish- urple

Figure 10.+.' ()* color system

0.8

/0l


11/23

10. olor 4epresentation - olor :ppearance 'ystem

he unsell color system is a typical color appearance system, in which color isrepresented with hue "%, saturation"'% and intensity "!% as a psychological response. ue is

composed of the five basic color; red "4%, yellow "3%, green")%, blue ">% and purple "% which

are located along a hue ring with intervals of C( degrees as shown in #igure 1$.6.1.

!ntermediate colors between the above five basic colors; 34, )3, >), > and 4 are locatedin between each other. #inally each hue is divided into ten but actually four. #or example 14,

B4, 1O4, 134, B34, 1O34, 13, ...... are a series along the hue ring.

!ntensity is an index of brightness with 11 ran?s from $ "dar?% to 1$ "light%. 'aturation is an

index of pureness ranging from $ to 16 depending on the hue and intensity. olor in the

unsell color system is identified as a combination of hue, intensity I saturation, for exampleB4= I 1$, which means B4 "hue%, = "intensity% and 1$ "saturation%.

#igure 1$.6.( shows a three dimensional color solid as called the unsellFs solid, with the

=$ panels with color samples of intensity and saturation with respect to the hue. unsell colorsamples are available in the commercial mar?et. :ny user can identity arbitrary colors by

comparison with the unsellFs color samples. sychologically defined '! has been correlated

with physically defined 4)> or 3xy as mentioned in 1$.B. herefore conversion between

4)> and '! can be made mathematically. !n the case of a color display using a digitalimage processing device, the 4)> signal has to be input, though color control is much easier

using '! indices. #igure 1$.6.+ shows the relationship between 4)> space and '! space.

he following are conversions from 4)> to '!, and from '! to 4)>. he ranges of

4,),>,',! are J$,1K , the range of is J$,(nK."1% from 4)> to '!

!0 ax. "4,),>%1% !0 $ ; ' 0 $, 0 indeterminalle

(% !HL $

' 0 "!-i%M1, where i 0min. "4, ), >N

/et r 0 "1-4% I "!-i%, g 0"!-)% I "!-i%, b 0 "1->% I "!-i%, then

if 4 = ! = "b-g% n I +

if

if

)0!

> 0 !

0"(9r-b%nl+

0 "=9g-r% 1t M+"(% from '! to 4)>

1% ' 0 $ ; 4 0 ) 0 > 0 ! regardless of value of

$) s* = +!1t h 0 floor"F% !f 0 2n, then 0 $

"floor "x% the function of getting the truncated value of x%

0 !"l-'%, O0 !"1-' "F- h%N, 0 ! "1-'"1-F9h%N, then

h 0 $ 4 0 !, ) 0 , > 0

h 0 1 4 0 , ) 0 !, > 0

h 0 ( 4 0 , ) 0 !, > 0

h 0 + 4 0 , ) 0 , > 0 !

h 0 = 4 0 !, ) 0 , > = O

h 0 B 4 0 !, ) 0 , > 0 O


12/23

7hite

Figure 1.2.$ 3hree dimensional color solid4 5unsells solid )

6

* y

c

Figure 1.2.! &elationshi7 between

&*% s7ace and 86s7ace

>lac?

Figure 1.2.15unsell color system


13/23

-1P8

10. 2perations et!een Images

Operations between multi-spectral images or multi-date images are very useful for image

enhancement and feature extraction.

Operations between images include two techniques; arithmetic operation and logical

operation.

a. 3rithmetic 2perations:ddition, subtraction, multiplication, division and their combinations, can be applied for

many purposes, including noise elimination. :s the results of the operation can sometimes

negative or small values between $ and 1, they should be ad5usted to a range, usually in

eight bits or $ to (BB for image display.

ypical operations are ratioing, for geological feature extraction, and normali*ed difference

vegetation index, for vegetation monitoring with AO:: :Q44 data or other visible near

infrared sensors.

"1% 4atioing4atio0 iX9

4atioing may be useful for geological feature extraction. 'uch ratioing can be applied to

multi-temporal thermal infrared data for extraction of thermal inertia.

($) 4ormali*ed ifference 5egetation lndex645I/

A


14/23

hannel Ao.1

4ationing

hannel Ao.(

A


15/23

-($$

10.8 Principal Component 3nalysis

Principal component analysis is used to reduce the dimensions of measured variables

"p dimension% to the representative principal components "m dimension , mRp % with a linear

combination of the measured variables. rincipal component analysis is applied to multi-band

data or multi-temporal data in order to reduce it to two or three principal components. !t should

be noted that in fact p components are formed, but the higher components of order greater than

m, usually contain noise and can be discarded.

/et the measured p dimensional variables be @xiN i 0 l,p, the principal components@*? N

? 0 1, m can be expressed as the linear combination as follows.

he coefficients "a1?- ap? % are determined under the following constrains.

(1) L ai?* 0 1

"(% Qariance a *? ( should be maximum

"+% *? and * ?9l should be independent of each other

he solution of the above problem can be obtained by determining the unique values and

the unique vectors which correspond to the variance and vector of the principal components

respectively.

he unique value represents the contribution ratio which indicates how much percentage

the principal component represents of the total tendency of the variables. he accumulativecontriution ratio percentage all the principal components represent of the total tendency of

the variables. Ssing an accumulative contribution ratio of 8$- P$ percent, will indicate how

many principal components should be adopted to effectively represent the ma5or variations in

the image data.

)raphically spea?ing, the first principal component for example in the case of two

dimensional variables "see #igure 1$.8.1% will be the principal axis which gives the maximum

variance. he principal component analysis can be used for the following applications.

(1) &ffective classification of land use with multi-band data

"(% olor representation or visual interpretation with multi-band data

"+% hange detection with multi-temporal data

!n the case of multi-band data with more than four bands, all bands cannot be assigned to

4, ) or> at the same time. owever the first three principal components can represent up to

five spectral variables with little information loss.

#igure 1$.8.( show the principal components and their color composite of /andsat

"6 bands%. )enerally the first principal component corresponds to the total radiance

"brightness%, while the second principal component represents the vegetation activity

"greenness%.


16/23

(

#

Figure 1.:.1 ;xam7le of 7rinci7al com7onent analysis

Figure 1.:.$ 7rinci7al com7onents andtheir color com7osite of ?83 35


17/23

10.9 &patial Filtering

&patial filtering is used to obtain enhanced images or improved images by applying,

filter function or filter operators in the domain of the image space "x,y% or spatial frequency

"x,h%. 'patial filtering in the domain of image space aims at image enhancement with so-called

enhancement filters, while in the domain of spatial frequency it aims at reconstruction with so

called reconstruction filters.

a. #iltering in the


18/23

Fig.10.,.1 Examples of spatial filters of + + 1in"o1

= -$ $ %= "finite

= -1 T%= "finite

-1 1 differences%

-1 A or -1 8 -1

-1 B or -1MP 8MP -1MP

-1 -1MP -1MP -1MP

a% original image b% 'obel c% /aplacian

d% smoothing e% median t) high pass

Figure 10.,.2 Image en!ancement it! use of 'x' operators

':!:/ #!/&:' ++ CD;&3C& #&'

'obel

6 : 6 9 6 > 6 or 3:(9>( where,

J-! $ -!K J-1 -( -1K-1 1 1 $ 1

gradient

differences%

reneit

6 : 6 9 6 > 6 or 3:(9>( where,

E-10

E-6-1 -K gradient

/aplacianJ$ -1

-

lJ !

-1

-1K -1 -1 -1 -1

differentia !

smoothingJ1MP 1MP 1MPK J $ 1MB $K

1MP 1MP 1MP or 1MB 1MB 1MB

1MP 1MP 1MP $ 1MB $

median 4eplaced with median of + X + window smoothed image

high-passJ $ -1

-n [-1/9 -1/9 -1/9] edge-enhancementsharpening

J 1MP -8MP 1MPK-8MP +CMP -8MP

1MP -8MP 1MP

clear image


19/23

10.10 ;exture 3nalysis

exture is a combination of repeated patterns with a regular frequency. !n visual

interpretation texture has several types, for example, smooth, fine, coarse etc., which are often

used in the classification of forest types. ;exture analysis is defined as the classification or

segmentation of textural features with respect to the shape of a small element, density and

direction of regularity.

#igure 1$.1$.1 "a% shows two different textures of density, while #igure 1$.1$.1 "b%

shows two different textures with respect to the shape of the elements.

!n the case of digital image, it is difficult to treat the texture mathematically because

texture cannot be standardi*ed quantitatively and the data volume is so huge.

owever texture analysis has been made with statistical features which are combined with

spectral data for improving land cover classification. ower spectrum analysis is another form

of textural analysis in which direction and wavelength or frequency can be determined forregular patterns of, for example, sea waves and sand waves in the desert.

a. Sse of 'tatistical #eatures

he following statistical values of an n x n window can be used as textural information

(1) )ray level histogram

"(% Qariance- co-variance matrix

"+% 4un-length matrix

hese values are used for classification together with the spectral data.#igure 1$.1$.( "a%shows the land cover classification using only spectral data while #igure 1$.1$.( "b% shows

the result of classification with spectral data as well as textural information. he result

shows a better classification for the urban area which has a higher frequency and variance of

image density.

b. :nalysis using ower 'pectrum

ower spectrum analysis is useful for those images which have regular wave patterns with a

constant interval, such as glitter image of the sea surface or wave patterns of sand dunes.

#ourier transformation is applied to determine the power spectrum which gives the

frequency and direction of the pattern.


20/23

-($B

U

)

U U U

U "a% different density

$? $ $

$ ?$

$ ? ? $ $

$ $ ? $$ $ $

$ $ $ $ $ ? $ $ $

"b% shape of elements

Figure 10.10.1 Texture analysis

a% only with spectral data b% with spectral data as well as textural

information

Figure 10.10.2 Classification it! use of texture analysis

U

$

$


21/23

10.11 !mage orrelation

!mage correlation is a technique by which the con5ugate point of a slave image "right%

corresponding to the master image "left% will be searched for the maximum correlation

coefficient. !mage correlation is applied to stereo images for ecause the above two correlations show almost no difference, the first correlation is

preferred to save computing time.

he si*e of the window should be selected depending on the image resolution and feature

si*e. #or example, B x B to P x P windows might be selected for 'O stereo images, while Px P to (1 x (1 would be better used for digiti*ed aerial photographs.

7hen the con5ugate points of stereo images are determined, the corresponding digital

elevation can be computed using collinearity equations based on photogrammetric theory.

#igure 1$.11.( shows the con5ugate points as white dots in a pair of 'O stereo

images, which were automatically recogni*ed by image correlation techniques.


22/23

hotographing direction hotographing directionof left image of right image

*

ryob5ect

left image right image

-----

7indow image

I

'earch using window

Ghotographing direction hotographing direction

t


23/23

Figure 10.11.1 conugate point Searc! 3 5 an" 467

on stereo image

Figure 10.11.2 utomatically recogni8e" conugate point

y image correlation tec!ni9ue

bright_inversion

Documents

Transcript of bright_inversion