Unit 2

Image Quality and Information content

Introduction
• Several factors affect image quality and information content of biomedical images

• Designing image processing techniques to remove the degradation and/or

improve image quality requires
• Good understanding of factors affecting quality
• Appropriate characterization of the associated loss in image quality

• Inherent problem in characterizing image quality: judged by human observers in a

subjective manner

• To quantify the notion of image quality is a difficult proposition

• The nature of the information conveyed by an image is difficult to quantify due to
its multifaceted characteristics in terms of
• Statistical
• Structural
• Perceptual
• Semantic
• Diagnostic connotations
Difficulties in Image Acquisition and Analysis
[1] Accessibility of the organ of interest
• Several organs of interest in imaging based investigation are situated well within the
body, encased in protective and difficult to access regions
• The brain is protected by the skull
• Prostate is situated at the base of the bladder near the pelvic outlet

• Several limitations are encountered in imaging such organs special imaging devices and
image processing techniques are required to facilitate their visualization

• Visualization of the arteries in the brain requires the injection of an Xray contrast agent
and the subtraction of a reference image

• Despite the use of such special devices and techniques, images obtained in some
applications tend to be affected by severe artifacts
[2] Variability of information
• Biological systems exhibit great ranges of inherent variability within their
different categories

• The intrinsic and natural variability presented by biological entities within a given
class far exceeds the variability that we may observe in engineering, physical, and
manufactured samples

• The distinction between a normal pattern and an abnormal pattern is often

clouded by overlap between the ranges of the features or variables used to
characterize the two categories; the problem is compounded when multiple
abnormalities need to be considered

• Imaging conditions and parameters could cause ambiguities due to the effects of
subject positioning and projection
• Most malignant breast tumors are irregular and spiculated in shape, whereas
benign masses are smooth and round/oval

• However, some malignant tumors may present smooth shapes, and some benign
masses may have rough shapes

• A tumor may present a rough appearance in one view or projection, but a

smoother profile in another

• The notion of shape roughness is nonspecific and openended

• Overlapping patterns caused by ligaments, ducts, and breast tissue in other

planes could also affect the appearance of tumors and masses in projection
images
• The use of multiple views and spot magnification imaging could help resolve
some of these ambiguities, but at the cost of additional radiation dose to the
subject
[3] Physiological artifacts and interference
• Physiological systems are dynamic and active

• Some activities, such as breathing, may be suspended voluntarily by an adult for

brief periods of time to permit improved imaging

• However, cardiac activity, blood circulation, and peristaltic movement are not
under one’s volitional control

• The rhythmic contractile activity of the heart poses challenges in imaging of the
heart

• The pulsatile movement of blood through the brain causes slight movements of
the brain that could cause artifacts in an giographic imaging
• Dark shadows may appear in ultrasound images next to bony regions due to
significant attenuation of the investigating beam and hence the lack of echoes
from tissues beyond the bony regions along the path of beam propagation

• An analyst should pay attention to potential physiological artifacts when

interpreting biomedical images

• Special techniques have been developed to overcome some of the limitations

mentioned above in cardiac imaging

• Electronic steering of the Xray beam has been employed to reduce the scanning
time required for CT projection data acquisition in order to permit imaging of the
heart
[4] Energy limitations

• In Xray mammography, considering the fact that the organ imaged is mainly
composed of soft tissues, a low kVp would be desired in order to maximize image
contrast

• However, low-energy Xray photons are absorbed more readily than high-energy
photons by the skin and breast tissues, thereby increasing the radiation dose to
the patient

• A compromise is required between these two considerations

[5] Patient safety:

• The protection of the subject in a study from electrical shock, radiation hazard,
and other potentially dangerous conditions is an unquestionable requirement of
paramount importance

• Most organizations require ethical approval by specialized committees for

experimental procedures involving human or animal subjects, with the aim of
• minimizing the risk and discomfort to the subject, and
• maximizing the benefits to both the subjects and the investigator.

• The relative levels of potential risks involved should be assessed when a choice is
available between various procedures, and analyzed against their relative
benefits
Characterization of Image Quality

• Images are complex sources of several items of information

• Many measures available to represent quantitatively several attributes of images

related to impressions of quality.
Changes in measures related to quality may be analyzed for:

• Comparison of images generated by different imaging systems;

• Comparison of images obtained using different imaging
• Parameter settings of a given system;
• Comparison of the results of image enhancement algorithms;
• Assessment of the effect of the passage of an image through a
• Transmission channel or medium; and
• Assessment of images compressed by different data
• Compression techniques at different rates of loss
Digitization of Images

• The representation of natural scenes and objects as digital images for

processing using computers requires two steps:
• sampling, and
• quantization.

• Both of these steps could potentially cause loss of quality and

introduce artifacts.
Sampling
• Sampling is the process of representing a continuous-time or continuous-space
signal on a discrete grid, with samples that are separated by (usually) uniform
intervals

• A band-limited signal with the frequency of its fastest component being fm Hz

may be represented without loss by its samples obtained at the Nyquist rate of fs
= 2 fm Hz.

• Sampling may be modeled as the multiplication of the given analog signal with a
periodic train of impulses.

• The multiplication of two signals in the time domain corresponds to the

convolution of their Fourier spectra.
The Fourier transform of a periodic train of impulses is another
periodic train of impulses with a period that is equal to the inverse
of the period in the time domain (that is, ls Hz).

Therefore, the Fourier spectrum of the sampled signal is periodic,

with a period equal to ls Hz.

A sampled signal has infinite bandwidth; however, the sampled

signal contains distinct or unique frequency components only up
to lrn = ± ls/2 Hz.

If the signal as above is sampled at a rate lower than ls Hz, an

error known as ahasing occurs, where the frequency components
above ls/2 Hz appear at lower frequencies.

It then becomes impossible to recover the original signal fron1 its

sampled version.
If sampled at a rate of at least f s == 2 f m Hz, the original signal
may be recovered from its sampled version by lowpass filtering
and extracting the base-band component over the band ±fm Hz.

If an ideal (rectangular) lowpass filter were to be used, the

equivalent operation in the time domain would be convolution
with a sine function (which is of infinite duration).

This operation is kno\iVn as ;nterpolation.

Figure 2.1 illustrates the loss of quality associated with sampling
an image at lower and lower numbers of pixels.

.. . r::- -~.,~
~~~
tJII!
~ \~,
.,. I
·' \'
-~
-~_~•"V''
';_,.
-.-=-
~- .•-.
-
'
~ •
.
--~
L
-~
....,,

(a) (b)

(c) (d)

Figure 2.1: Effect of sampling on the appearance and quality of an image: (a) 225 x 250 pixels; (b) 112 x 125
pixels; (c) 56 x 62 pixels; and (d) 28 x 31 pi,xels. All four images have 256 gray levels at 8 bits per pixel.
Quantization

Quantization is the process of representing the values of a

sampled signal or image using a finite set of allowed values.

Using n bits per sample and positive integers only, there exist
2n possible quanhzed levels, spanning the range [O, 2n - 1].

If n == 8 bits are used to represent each pixel, there can exist

256 values or gray levels in the range [O, 255].
It is necessary to map appropriately the range of variation of the
given analog signal to the input dynamic range of the quantizer.

The decision levels of the quantizer should be optimized in

accordance with the probability density function (PDF) of the
original signal or image.

The Lloyd-Max quantization procedure is used to optimize a

quantizer
Lloyd-Max quantization:

Let p( r) represent the PDF of the amplitude or gray levels in the

given image, with the values of the continuous or analog variable
r varying within the range [r min"' r max] .

Let the range [rmin"' rmax] be divided into L parts demarcated by

the decision levels Ro, R1, R2, ... , RL, with Ro = rmin and
RL = rma..,-x; see Figure 2.2.

L output levels of the quantizer represent the values

Let the
Qo, Q1, Q2, · · · ~ QL-1-
The mean-squared error (MSE) in representing the analog signal
by its quantized values is given by

L-1 R1 2
L fR, +
1
(r - Qz) p(r) dr. (2.1)
l=O

Q Q Q Q Q Q Quantizer
0 1 2 3 L-2 L-7 output levels
I 1
R
0
= rmin R
1
R
2
R
3
R
4
R
L-1
R
L
r
max
Decision levels

gray l.evel. r

Figure 2.2: Quantization of an image gray-level signal r with a Gaussian (solid line) or uniform (dashed line)
PDF. The quantizer output levels are indicated by Qi and the decision levels represented by Fq.
A classical result indicates that the output level Q z should lie at
the centroid of the part of the PDF between the decision levels
Rz and Rz+ 1 > given by
If{z+i r p( r) dr
Qz = R 1
, (2.2)
IR/+ p( r) dr
which reduces to
Rz + Rz+1
(2.3)
Qz
2
if the PDF is uniform.

The decision levels are then given by

Rz = Qz-1 + Qz . (2.4)
?.
.,_,

The use of an inadequate number of quantized gray levels leads to

false contours and poor representation of image intensities.
Table 2.1: Relationships Between Tissue Type, Tissue Density, X-ray Attenuation Coefficient, Hounsfield Units
(JIU), Optical Density (OD), and Gray Level. The X-ray Attenuation Coefficient was :vleasured at a Photon
Energy of 103.2 keV.

Tissue Density X-ray Hounsfield Optical Gray level Appearance

type gm/cm3 atten. (cm- 1) units density (brightness) in image

lung < 0.001 lower low high low dark

[-700, -800]

liver 1.2 0.18 medium medium medium gray

(50, 70]

bone 1.9 higher high low high white

[+800, +1, 000]
Array and Matrix Representation of Images
Images commonly represented as 2D functions of space: f(x, y).

A digital image f ( rn, n) may be interpreted as a discretized

version o f f ( X, y) in a 2D array~ or as a matrix.

Notational differences between the representation of an image as

a function of space and as a matrix: source of confi1sion _,

An J\I X JV matrix has J1I rows and JV columns;

its height is J1I and width is J\T;

numbering of the elements starts with (1, 1) at the top-left corner

and ends with (J1I, JV) at the lower-right corner of the image.
A function of space f ( x, y) that has been converted into a digital
representation f ( rn~ n) is typically placed in the first quadrant in
the Cartesian coordinate system.

Then:> an JVJ X JV will have a width of J\J and height of J\T;

indexing of the elements starts with (0, 0) at the origin at the

bottom-left corner and ends with (Jvl - 1, N - 1) at the
upper-right corner of the image.

The size of a matrix is expressed as rows X colurnns:>

the size of an in1age is usually expressed as width X height.

 column n =1 2 3 4

row
2 f(0,2) f(l,2) f(2,2) f (3, 2) f(l,1) f(l,2) f(l,3) f(l,4)
m =1

1 f(0,1) f(l,1) f(2,1) f (3, 1) 2 f (2, 1) f(2,2) f(2,3) f(2,4)

y =0 f ( 0, 0) f(l,0) f(2,0) f (3, 0) 3 f (3, 1) f(3,2) f(3,3) f(3,4)

X =0 1 2 3

f(x, y) as a 4x3 function of f(m, n) as a 3x4 matrix

space in the first quadrant (in the fourth quadrant)

Figure 2.4: Array and matrix representation of an image.

Optical Density

• The value of a picture element or cell—commonly known as a pixel,

or occasionally as a pel — may be expressed in terms of a physical
attribute such as temperature, density, or X-ray attenuation
coefficient; the intensity of light reflected from the body at the
location corresponding to the pixel; or the transmittance at the
corresponding location on a film rendition of the image
• The OD at a spot on a film is defined as

Ii]
OD - log 10 [ Io

• where Ii is the intensity of the light input and Io is the intensity of the light
transmitted through the film at the spot of interest

A perfectly clear spot will trans1nit all of the light that is input c::::::c~::::::::l
and will haYe OD= 0;
l 1

J.
•I•o-==i-- film with image
(transparency)

.l

light

a dark spot that reduces the intensity of the input light by a factor Measurement of the optical density
at a spot on a film or transparency
of 1, 000 will haYe OD = 3. using a laser microdensitometer.

X-ray filtns: OD~ 0 to OD~ 3.5.

Dynamic Range

• The dynamic range of an image is usually expressed as the difference

between the maximum and minimum values present in the image

• X-ray films for mammography typically possess a dynamic range of 0 –

3.5 OD

• Modern CRT monitors provide dynamic range of the order of 0 – 600

cd/m2 in luminance or 1:1000 in sampled gray levels

cd - The candela is the SI unit of luminous intensity. The candela is used to measure the visual intensity of
light sources, like light bulbs or the bulbs in torches. It is the only SI base unit based on human perception
 • Figure below compares the characteristic curves of two devices

Device A has a larger slope or “gamma”

3.0 Saturation than Device B, and hence can provide
--------------------------
Shoulder /,,.- higher contrast
/
/
I
I
I
2.0 Device A ,
I
Device B Device B has a larger latitude or
I
Optical I
) breadth of exposure and optical
I
density
I
( density over which it can operate than
) ,
I , Device A
1.0 1·
'/
- ,I

Toe
/
/
Plots of film density versus the log of X-
~--
,I
.,.. . / Back round level
(base, fog, noise)
ray exposure are known as Hurter-
Driffield or HD curves
log (exposure)
• Air in the lungs and bowels, as well as fat in various organs including
the breast, tend to extend the dynamic range of images toward the
lower end of the density scale

• Bone, calcifications in the breast and in tumors, as well as metallic

implants such as screws in bones and surgical clips contribute to high
density areas in images

• Mammograms: dynamic range of 0 − 3.5 OD

• CT images: dynamic range of −1, 000 to +1, 000 HU

Hounsfield units (HU) are a dimensionless unit universally used in computed tomography (CT) scanning to
express CT numbers in a standardized and convenient form. Hounsfield units are obtained from a linear
transformation of the measured attenuation coefficients
Contrast
• Contrast is defined in a few different ways, but is essentially the difference
between the parameter imaged in a region of interest (ROI) and that in a suitably
defined background
• If the image parameter is expressed in OD contrast is defined as
Coo = f oo - boo
where foD and boD represent the foregrom1d ROI and
background OD. respectively.

Figure 2.7: Illustration of I.he notion of contrast. comparing a foreground region / with its badtground b.
vVhen the in1age para1neter has not been nonualized. the n1.eastu·e
of contrast will require nonnalization.

If. for exa111ple. f and b represent the a,-erage light intensities

e1nitted or reflected from the foreground ROI and the
background. respectiYely. contrast 1nay be defined as
f-b
C = f + bi (2.7)

or as

f-b
C1 = -b-. (2.8)
Due to the use of a reference background. the measures defined
aboYe are often referred to as s;,nulraneous contrast.
• Example: The two squares in Figure are of the same value (130 in the Scale 0 –
255), but are placed on two different background regions of value 150 on the left
and 50 on the right

• The lighter background on the left makes the inner square region appear darker
than the corresponding inner square on the right
• This effect could be explained by the measure of simultaneous contrast: the contrast of the inner
square on the left, using the definition in Equation 2.8 is

_ 130 150 _
Cl - - 0.1333,
150
whereas that for the inner square on the right is
130 50
Cr - - tl.6.
50
• The values of Cl and Cr using the definition in Equation 2.7 are, respectively, -0.0714 and +0.444
• The advantage of this formulation is that the values of contrast are limited to the range [-1,1]

• The negative contrast value for the inner square on the left indicates that it is darker than the
background, whereas it is the opposite for that on the right
Just noticeable difference
• The concept of just-noticeable difference (JND) is important in analyzing contrast,
visibility and the quality of medical images
JND is determined as follows
• For a given background level b as in Equation 2.8, the value of an object in the
foreground f is increased gradually from the same level as b to a level when the
object is just perceived
• The value (f – b)/b at the level of minimal perception of the object is the JND for
the background level b
• The experiment should, ideally, be repeated many times for the same observer,
and also repeated for several observers
• Experiments have shown that the JND is almost constant, at approximately 0.02
or 2% over a wide range of background intensity; this is known as Weber’s law
 Example: The five bars in Figure 2.9 have intensity values of (from left to
right) 155, 175, 195, 215, and 235. The bars are placed on a background of
150. The contrast of the first bar (to the left), according to Equation 2.8, is

C _ 155 - 150 _ +0.0 33 _ (2.11)

l 150
This contrast value is slightly greater than the nominal .JND; the object should
be barely perceptible to most observers. The contrast values of the remaining
four bars are more than adequate for dear perception.

1111
Example: A calcification that appears again.st fat and lo~•-density
tissue may possess high. contrast and be easily Yisible.

A sini.ilar calcification against a background of high-density

breast tissue. or a calcification ~rithin a high-density nrmor. could
possess low contrast. and be difficult to detect.

Figure 2.10: Part of a mammogram with several calcifications ~iatcd with malignant b~t disca1:£. The
density of the background affects the contrast and visibility of lhe calcificalions. The image h~ 768 x512 pixe~
at al'CSl]ution of 62 µm; the true \\idth of lhe image~ about ~2 mm.
Histogram
• The dynamic range of the gray levels in an image provides global information on
the extent or spread of intensity levels across the image

• However, the dynamic range does not provide any information on the existence
of intermediate gray levels in the image

• The histogram of an image provides information on the spread of gray levels over
the complete dynamic range of the image across all pixels in the image
Histogram Processing
• The histogram of a digital image with gray levels in the range [0, L-1] is a
discrete function
h(rk)=nk,
where
rk is the kth gray level
nk is the number of pixels in the image having gray level rk
• In practice, a histogram is normalized by dividing each of its values by the
total number of pixels in the image, denoted by n
• A normalized histogram is given by
p(rk)=nk / n,
for k=0, 1,…… ,L-1
• p(rk) gives an estimate of the probability of occurrence of gray
level rk

• Histograms are the basis for numerous spatial domain processing

techniques

• Histogram manipulation can be used effectively for image

enhancement

• Histograms provide useful image statistics

The histogram of an image shows us the distribution
of grey levels in the image. It is useful in image
processing, especially in enhancement and
segmentation.
1 2 3 4 5 6

I
2

6
3

6
4

4
5

6
6

6
cP 5

5/36
4

4/36
5

5/36
6

6/36
2

2/36
14

14/36

3 4 5 6 6 6

I 4 6 6 2 3

I 3 6 4 6 6

42
 300(),------~------~------~-------,

200

43
 Properties of Histogram
♦ The histogram only shows the distribution of grey levels in the
image, and it doesn't include the location information of pixels.

♦ One iinage has its corresponding histogram, but different rmages

may have the same histograms.

(a) (b)

44
lniage Histogran,s

X 10
.
..
3.5

2.5

1.5

0.5

0 50 100 150 200 250

It is a baby in the cradle~

Histogran1 infor1nation reveals that image is under-exposed

45
Image Histograms

7000

6000

5000

4000

3000

2000

1000

Over-exposed image

46
Groups of Histogrm:n

The horizontal axis

of each histogram Darlc image Bright image
plot corresponds to
gray level values rk

The vertical axis

corresponds to values JJ(r,,)

of p(ric) if the values

are normalized. Low-contrast Hig)1-contrast
image image

47
Grol1ps of Histogram

MlmH::ltl:c:i::i
■ h:C:im111c::i:
► lmli· =
·Uc:ll:Ci ~

Low contrast High contrast Dark image Bright image

48
• Information inherent in histograms also is quite useful in other
image processing applications, such as image compression and
segmentation

• Histograms are simple to calculate in software

• Histograms lend themselves to economic hardware

implementations
• Popular tool for real-time image processing
Role of histogram processing in image enhancement

Dark image
For dark image components
of the histogram are
concentrated on the low
(dark) side of the gray scale

Bright image
The components of the
histogram of the bright image
are biased toward the high
side of the gray scale
 Low-cootrasl in,age
The components of the
histogram in the high-
contrast image cover a
broad range of the gray
scale

High-contrast image The distribution of pixels

is not too far from
uniform

An image with low contrast has a histogram that will be narrow and will be centered
toward the middle of the gray scale
For a monochrome image this implies a dull, washed-out gray look
 LS

t
ti.
b

I,
z

o.s

....,so_ _•
0 0....__ _ _
100 ,so 200 2SO
Gl>ylevel

Figure 2.11: Histq;ram of the imai;coflhc ,-mtricular myocytc in Figure 1.3. The size of the llMllC is 180 x 480 •
230,400 pixels. E:r1lropy // 4.96 bola.

• It is seen that most of the pixels in the image lie in the narrow range of 70 – 150 out of the available range
of 0 – 255
• The effective dynamic range of the image may be taken to be 70 – 150 rather than 0 – 255
• This agrees with the dull and low-contrast appearance of the image
• The full available range of gray levels has not been utilized in the image which
could be due to poor lighting and image acquisition conditions, or due to the
nature of the object being imaged

• The gray level of the large blank background in the image is in the range 80 -90;
the peak in the histogram corresponds to the general background range

• The relatively bright areas of the myocyte itself have gray levels in the range 100 –
130

• The histogram of the myocyte image is almost unimodal; that is, it has only one
major peak

• The peak happens to represent the background in the image rather than the
object of interest
 ,,.,.

•~ e '
•-•r,·:~,:,··~~--.
:-t>.
• . • - ·o~
~ :a ~I
·~~~1~..
•
I .--. i i
••'5
o.il,,a~
••o. ··--o
•• •:~ ~
•• .'¥':..,
~•J: l~ •~Ffofl
,• •F. !ti eS; 1
I _lit..,., ~ \~

~.. •.at°'~
~ , a
t: t.·-·-~
,_

'"""
t
! tCDl

! .,.
2

GO

:z,c

!Ill
010!-_ _. . . , . _ _. . . .

Figure 2.12: (a) Histogram of the image of the collagen fibers in rigure 1.5 (b); // 7.0 bit.s. (b) Histogram of (b)
the image after the application of the 3 x 3 mean filter and rounding the results to int.cgcrs; // 7.1 bil.s.
 0.025

0.02

8
~ij 0.015
'II
l:-
D
.8
£ 0.01

0.005

0
0 50 100 150 200 250
Gray level

Figure 2.14: Normalized histogram of the nuunm0£ram in Piguro 2.13. Entropy II 6.92 bi.ll.

F'igure 2.13: Part of a mammograrn with a malignanl lwnor (the rclalivcly bright region along lhc upper-left
odgc of lhc image). The size of lhc image is 700 x 700 490,000 pixels. The pixel resolution of 62 µm; the widlh
of Lhc image is about 44 mm. Image courtesy of Foothills Hospilal, Calgary.
 -

(a)

j•
i •
!
Figure 2.15: CT ima~ of apatient with ncuroblastoma. Only one scetional image out of atot.al of 75 llll8o"tS in i.
the study is shown. The sil.c of the image is 512 x512 262.1~1 pixels. The tumor, which appears as alarge
circular rq;ion oo thr left-hand side of the ima~1 includes calcified t~cs that appear as bright regions. The
/JU ran~ of I 200.400 has been linearly mapped lo the dL5play range of!O, 2551; s:c BBl figurcs 2.16 and 4.4.
Image court~ of AIberia Childrtn's lla;pilal. Calgary.
•100 o

-- t:x>

(b)
XO X10 c,o

l'igurc 2.16: (a) Histogram of lhc CT iMion image in Figure 2.15. (b) llislogram of the cnlirc CT sludy or the
I
palicnl. \\ilh 75 ll:Clional im,~. The histograms arc ~~rlaycd for the range 1/U 2Wa400 only. I 12 =
 Entropy

The distribution of gray levels over the full available range is represented by the
histogram

The histogram provides quantitative information on the probability of occurrence of

each gray level in the image

However, it is often desirable to express, in a single quantity, the manner in which the
values of a histogram or PDF vary over the full available range

Entropy is a statistical measure of information that is commonly used for this purpose
Entt·opy is a stahstical n1eas11re of infor1nation.

Pixels in an i1nage considered to be sy1nbols produced by a

discrete infortnation sotu·ce '1Vith the gray leYels as its states.

Consider the occurrence of L gray levels in an image.

'1vith the probability of occurrence of the l th gray leYel being

p(l), l = 0, l, 2, ... , L - l.

Gray leYel of a pixel: a rando1n Yariable.

A measure of information conveyed by an event, a pixel or a gray level may be related to the
statistical uncertainty of the event giving rise to the information rather than the semantic or
structural content of the signal or image

A 1n.easure of information h(p) should be a function of p( l).

satisfying the follo"'1ing criteria:

• h (p) should be continuous for O < p < 1.

• h(p) CX) for p = 0.
• h(p) 0 for p = 1.
• h(p2) > h(p1) ifp2 < Pl·
• If tvvo statistically it.1.dependent itnage processes (or pixels)
f and g are considered. the joint information of the tv. 0 1

sources is the su1n of theit- individual meas1.ues of information:

h 1 ,9 = h 1 + h 9 .

These requit·e1nents are in.et by h(p) - log(p).

When a source generates a tnunber of gray le-vels with different
probabilities. a rneasure of ayerage infonnation or entropy is the
expected value of infon.nation in each possible level:
L-1
H = E p(l) h[p(l)]. (2.17)
l=O

Using - log2 in place of h. vve obtain

L-1
H = - lEo p(l) log2 [p(l)] bits. (2.18)

Because the gray leYels ai-e considered as indiYidual entities in

this definition. that is. no neighboring elements are taken into
account. the result is knovvn as the ::erotlz-order entropy.
Differentiating the function in Equation 2.18 with respect top( l).
it can be shovn1 that the n1axinu11n possible entropy occurs when
all the gray leYels occur with the san1e probability (equal to l):
·when the various gray le,·els are equally likely:

L-1 1 [ 11 . (2.19)
Hrna'<= - I: L log2 Lj = log2 L.
l=O

If the nu1nber of gray leYels in an in1age is 21(. H ma'< = K bits.

Maxinuun possible entropy of an i1nage ·with 8-bit pixels is

8 bits.
Entropy characterizes the stahstical infonnation content of a
source based upon the PDF of the constituent events. vvhich are
treated as rando1n variables.

The 111easure is not sensitiYe to the pictorial. strucnu·al. se1nantic.

or application-specific (diagnostic) info1u1ation in the i1nage.

Entropy does not account for the spatial distribution of the gray
leYels in a given i1nage.

Gives the lo\7\rer bound on the noise-free trans1nission rate and

storage capacity requiren1ents.
Properties of entropy:

• HP > 0, -w-ith HP= 0 only for p = 0 or p = 1:

• The joint information H(Pi,'P2,···,Pn) conYeyed by n events_ ,.vith
probabilities of occurrence Pl, P2, • • • , Pn· is goYerned by
H(Pt-P2·"'-Pn) < log(n).
'A-itb equality if and only if Pi = ~ for i = 1, 2, · · · , n.
■ Considering t\vo i1nages or sotu·ces f and g V\ ith PDFs p 1( l1)
1

and Pg( l2). where l1 and l2 represent gray leYels in the range
[O, L - l]. the average joint infonuation or joint entropy is

If the tvvo sources are statistically independent. the joint PDF

PJ,g(li, l2) reduces to PJ(l1) Pg(l2).
Joint entropy is goven1ed by the condition
HJ.g <HJ+ Hg.
\7\ ith equality iff f and g are statistically independent.
7
■ The conditional entropy of au i1nage f given that another
i111age g has been observed is

\:\•here Pf g(l1, l2) is the conditional PDF off giYen g.

Hflg == HJ. 9 - H 9 < H 1.
v\-ith equality iff f and g are statistically independent.
;Vote: The conditional PDF off giYen g is defined as

(2.22)
1 otherwise.
Higher-order entropy: The definition of the zeroth-order
entropy in Equation 2.18 assu1nes that the successiYe pixels
produced by the source are statistically independent.

While governed by the lin1it Hmax == K bits. the entropy of a

real-v.,orld i1nage could be considerably lov.,er:

neighboring pixels are not independent of one another.

It is desirable to consider sequences of pixels to esti1nate the tn1e

entropy or infonnation content of a given i1nage.
Let p( { ln}) represent the probability of occurrence of the
sequence { lo> l1, l2, • · • , ln} of gray leYels in the i1nage f.

n: 11u111ber of neighboring or additional ele1ne11ts considered. not

cotulting the initial or zeroth ele111ent.
The n th -order entropy off is defined as

1
I: p( { ln}) log2 [p( { ln} )] ; (2.23)
(n + 1) {ln}

E over all possible sequences { ln} v;ith ( n + 1) pixels.

Variations exist in the definition of higher-order entropy.

H n is a monotonically decreasing function of n. and approaches

the true entropy of the source as n -----* CX).
Mutual information:

Important in analysis of transn1ission of images oYer a

conununication systen1. as vvell as storage in and retrieval fro1n an
a1·chiYal syste1n. vvith potential loss of inforn1.ation.

Represents the inforn1ation receiYed or retrieYed:

HJ is the information input to the trans1nission or archiYal system

in the fonn of the i1nage f.

Hf 9 is the infonnation about f giYen that the received or

retrieved image g has been obse1Yed.
If g is completely cotTelated vvith f

Hf g = 0 and I fig = HJ

no loss or distortion in i111age transmission and reception.

If g is independent off

H1 g = HJ and I11 9 = 0

complete loss of infonnation in transn1ission or a1·chiYal.

Entropy and 1nun1al infonnation useful in the design and analysis

of i1nage archival. coding. and conununication systems.
Blur and Spread Functions
SeYeral con1po11ents of image acquisition syste1ns cause blurring
due to intrinsic and practical limitations.

The sitnplest Yisualization of blun·it1.g is proYided by using a

single. ideal point to represent the object being in1aged.

(a) (b)

Figure 2.17: (a) An ideal point source. (b) A Gaussian-shaped point spread function.
Mathematically. an ideal point is represented by the continuous
unit itnpulse function or the Dirac delta function 8 (x, y):

b(x, y) = { undefined at x = ~' y = 0 (2.25)

0 otherwise,

and

fx00 _ 00
Jy00 _ 00
8(x, y) dx dy = 1. (2.26)
The JD Dirac delta function:
-----
8(x) is defined in tenus of its action ·within an integral as

lb f (x) 8(x - Xo) dx = { f (xo) if a < ~o < b (2.27)

a O otherwise,

where f (x) is a function that is continuous at x 0 .

This is known as the sifting property.

The expression 1nay be extended to all x as

f(x) = f~-oo f(a) b(x - a) da. (2.28)

Resolving the arbitra1y signal f (x) into a weighted c01nbination

of 111utually 01thogonal delta functions.
Delta function in tenns of its integrated strength:

00
/ 00 8(x) dx = 1, (2.29)

vvith the conditions

(5(x) == {undefined at x == 0 (2.30)

0 otherwise.
Delta function in te11ns of its integrated strength:

00
/ 00 o(x) dx == 1, (2.29)

with the conditions

o(x) == {undefined at x == 0 2 30
0 otherwise. ( - )
The delta function is also defined as the liiniting condition of
several ordinaiy functions. one of which is

o(x)=lim_!_ exp(-lxEI). (2.31)

t:-0 2E

The delta function 1nay be visualized as the limit of a function

with a sha1p peak of undefined value, whose integral over the fbll
extent of the independent variable is maintained as unity while its
te1nporal or spatial extent is compressed toward zero.

The image obtained when the input is a point or m1.pulse function

is known as the impulse response or

point spread funcrion (PSF).

Assuming the iinaging systern to be linear and shift-invariant
(LSI) (or position-invariant or space-invariant), the itnage g(x, y)
of an object f(x, y) is given by the 2D convolution integral

g(x, y) = !:° _ Jfl° _

00 00 h(x - a, y - /3) f(a, /3) da df3

= J;o -oo J!J° -oo h(a, /3) J(x - a, y - /3) da d/3

= h(x, y) * f(x, y), (2.32)

where h(x, y) is the PSF. a and f3 are tempora1y vaiiables of

integration, and * represents 2D convolution.
Some examples of the cause ofbluning are:

■ Focal spot.
■ Thickness of screen or c1yst.al.
• Scattering.
Ideal
point
source J Finite
focal
spot

-x rays

Object
\-being
\ imaged
:' \ \
: \
' \

Umbra Umbra
Penumbra
(blur)

l<iguro 2.18: The effect of a finite focsl spot (X-my-genemling ponion of the target) on the sh9.f))ncss of the image
of an object.
Point, line, and edge spread functions:

It is often not possible to obtain an itnage of an ideal point.

However, it is possible to construct phanton1S to represent ideal

lines or edges.

An itnage obtained of line function is la1own as the

line spread funcrion (LSF) of the systetn.

A cross-section of an ideal straight line is a point (impulse).

The reconstruction of a cross-section of a line phanto1n provides

the PSF of the system.
 6 (x~ y) y f (x,y) y
1

Point-------- Line Edge

l,igurc 2.19: The relationship between point (impulse function), line, and edge (step) images. The height of each
function represents its strength.
A phantom representing an ideal edge may also be used.

A profile of the image of such a phantom across the ideal edge

provides the edge spread f11ncrion (ESF).

The de1ivative of the ESF gives the LSF of the system.

The PSF may be estimated :frorn the LSF.

Ideal (sharp) edge

f (b) -,-

Blurred or
unsharp edge
,
J
Intensity
,
J
f(x) ,
J
,J
f(a)

I I

x=a x=b
Distance x

r,'igure 2.20: Blurring or an idCAI sharp edge into an unsharp edge by an imaging system.
Mathe1natical relationships between the PSF, LSF, and ESF:

Consider integration of the 2D delta function along the x axis:

00
fl ( X, Y) == fx -oo <5 (X, Y) dx

== Jxoo - oo <5 (x) <5 (y) dx

== o(y) fx00 _ 00 <5(x) dx

== o(y). (2.33)
The last integral above is equal to unity.

The separability property of the 2D ilnpulse function as

o(x, y) == o(x) o(y) has been used above.

o(y) over the 2D (x, y) space is a line fimction on the x axis.

The output of an LSI system when the input is the line image
fz(x, y) = b(y). that is. the LSF hz(x, y), is

hz(x,y) = f~-oo f~-oo h(a,{3) fz(x-a,y-{3) dad{3

= f~-oo Jft=_ 00 h( a, {3) b(y - {3) da d{3

= f~-oo h(a, y) da

= fx~-oo h(x, y) dx. (2.34)

h( x, y) is the PSF: the LSF is the integral of the PSF.

Consider the Fou1ier transfonn of hl(x, y).

Hz(v) == Jy
00
_ 00 hz(y) exp(-j21rvy) dy

== Jy
00
_ 00 dy fx00 _ 00 dx h(x, y) exp[-j21r(ux + vy)]lu=O

== H(u, v)lu=O

== H(O, v), (2.35)

H(u, v) is the 2D Foulier transfonn of h(x, y):

the 111odulaho11 tra11sfer f1111ctio11 O\IJTF).

The Fou1ier transfo1m of the LSF gives the values of the Fourier
transfo1m of the PSF along a line in the 2D Folll'ier plane
(in this case. along the v axis).
Consider integrating the line function as follows:

fe(x, y) == !$ -oo fz(x, {3) d/3

== JJ -oo 8(/3) d{3. (2.36)

The resulting function has the property

1 if y > 0
Yx, fe(x, y) == (2.37)
10 if y < o,

which is an edge or unit step function parallel to the x axis.

Thus, the edge or step fhnction is obtained by integrating the

line function.
The ESF is given by

he(Y) == fJ,__ 00 ht(f3) d{3. (2.38)

Conversely. the LSF is the derivative of the ESF:

d
ht(Y) == dy he(y). (2.39)

Thus the ESF 1nay be used to obtain the LSF, which 1nay ftuther
be used to obtain the PSF and MfF.
Fro1n the Fourier slice theorem:

The Fourier transfonn of a profile of the LSF is equal to the radial

profile of the Folu-ier transfo1m of the PSF at the angle of
placernent of the line source.

If the ilnaging systern rnay be assumed to be isotropic iI1 the plane

of the liI1e source, a single radial profile is adequate to reconstn1ct
the complete 2D Fourier transfonn of the PSF.

An iI1verse 2D Fou1ier transfo1m provides the PSF.

Line solll'Ce prepared using a plastic tube of intenial radius
1 mm. filled with 1 mCi (milli Cmie) of 99mTc.

Figure 2.21: ~uclcsr medicine (planar) image of a line source obtained using a gQrnma carncr11. The size of the
image is 64 x &I pixels. with an effective width of 100 mm. The pixel siw is 1.56 mm.
 180

160

140

120

! 100
§
I
"'
80

60 I
I
I \
'

40 I
20

0 .___
0 10 20 30
J
___.._ __,__ _. . . . _ - = = - - ' - - - - - ' - -
40 50
Distance in nvn
60
\_
70 so 90 100

Figure 2.22: Sample profile (dot Led line) and averngcd profile (solid line) obtained from the image in Figure 2.21.
Either profile may be taken to represent. the LSF of the g11111m11 camera.

Full width at half the maximurn (FWHM): 0. 5 - 1. 7 cm.

Resolution
The spatial resolution of an ilnaging syste1n or an ilnage 1nay be
expressed in tenns of:

• The sainpling inte1val (in, for example, mm or µm).

■ The width of (a profile of) the PSF, usually FWI-Ilv1 (in mm).
■ The size of the laser spot used to obtain the digital image by
scanning an 01iginal fihn, or the size of the solid-state detector
used to obtain the digital image (in µm).
■ The smallest visible object or separation between objects in
the irnage (in mm or µm).
■ The finest grid patten1 that re1nains visible in the ilnage
(in lp/mm).
Typical resolution limits of a few itnagit1g syste1ns:

■ X-rayfihr1: 25 - 100 lp/mm.

■ screen-film co111bit1ation: 5 - 10 lp / mm:
111allllnography: up to 20 lp/mm.
■ CT: 0.7 lp/mm:
µCT: 50 lp/mm or 10 µm:
■ SPECT: < 0.1 lp/mm.
The Fourier Transform and Spectral Content

The Fou1ier transfo1m is a linear, reversible transforrn that maps

an itnage frotn the space dotnain to the frequency domain.

Conve11ing an itnage from the spatial to the frequency (Fourier)

domain helps in

■ assessing the spectral content.

• assessing the energy distlibution over frequency bands,
• designing filters to remove noise,
■ desi going filters to enhance the itnage,
• extracting ce11ait1 components that are better separated in the
frequency dotnain than in the space dotnain.
2D Fourier trausfonn of an image f (x, y) is denoted by F( u, v):

F(u, v) = fx~-oo Jy~-oo f(x, y) exp[-j 21r(ux+vy)] dx dy.

(2.40)

u. v: frequency in the horizontal and ve1tical directions.

exp[-j 21r (ux + vy)] (2.41)

= exp(-j 21rux) exp(-j 21rvy)

= [cos(21rux) - j sin(21rux)] [cos(21rvy) - j sin(21rvy)].

Irnages are typically :fhnctions of space.

Units of rneasurement in the i1nage do1nain:

rn, cm. rnrn. µrn. etc.

In the 2D Fou1ier do1nain. the unit of frequency is

cycles/rnrn. cycles/rn, rnrn- 1 , etc.

Frequency is also expressed as l p / rnrn.

If the distance to the viewer is taken into account. frequency

could be expressed in tenns of cycles/ degree of the visual
angle subtended at the viewer's eye.

The tulit Hertz is not used in 2D FoUiier analysis.

It is co1runon to use the discrete Fou1ier t:ransfonn (DFT) via the
fast Fourier transfonn (FFn algo1ithm.

2D DFTof a digital in1age J(m, n) of size M x N pixels:

1 AI-1 N-1 [ nf\l(mk

F(k,l) = MN I: I: J(m,n) exp -j21r M + NI .
m=O n=O \ )
(2.42)

For con1plete recove1y of J(m, n) fro1n F(k, l), the latter

should be c01nputed fork= 0, 1, ... , M - 1, and
l = 0, l, . . . , J\T - l, at the nllllllntun.
Then, the inverse transfonn gives back the 01iginal image with no
enor or loss of infonnation as

f (mk nl'\l
f (m, n) == AI-1
I:
k=D
N-1
I:
l=D
F(k, l) exp +j 21r M +NI ,
l )J
(2.43)

for m == 0, 1, ... , M - L and n == 0, 1, ... , N - 1.

This expression 1nay be inte1preted as resolving the given itnage

into a weighted sun1 of 1nutually 01thogonal exponential
(sinusoidal) basis functions.
 I J [It r 11111111 I 11 I 11111111 tr [ J. 11 11
•' t I I 1[ 1] 11111111 I 1111 I I t ' •

• r 1111111 111 r, I t[ r I 111IIlJ111.

11
' 11 I l I 111111 1 •. j I I I I I I I [ 111 I

, I I 111 I Ir , , , I I 11 I I Ir . r I I I l I I I 1,
1
'I! I I l I l 1 ' I[ I I I I 11 1I I I I I I I 11 1 I
11 I 11 r 1
1
, 1111 r,
1 1
d 11 I I I d 11 I I) 1
111111 11111 I I I I I] I I I I 111

I I I I Ir I I l 11 I I I JI I I [ ,1111, irllr,
I I
I 111 II II I' 11 I I I I ' 1111 1' I I I l I'
I 11 I I I I 11 I I IIII ti I 1. , r 11, 1II I1
I IlII I I 11' ' [ I 11 I IIII l I I I' l IlII

I] I 1 1 111, r 11 I r 11 I c 11 1. , I l 1. . r I le
III 'I I I' •I I I 1 I III I 11 1 [II I I l II
l<igurc 2.23: The firaL eight sine basis functions of the ID OFT: k = 0, l, 2, ... , 7 from top to bottom. Each
function was computed using 64 samples.
lligurc 2.24: The first 6.1 cosine basis functions of the 2D Dl-'T. Each function was computed using a 64 x 64
matrL'I:.
Fij,'llrc 2.25: The first 6-1 sine be.sis funeLions of the 2D DIT. Each funcLion was computed using e. 6.J x 6-t me.trix.
Zero padding for FFT: Pad the given i1nage with zeros or so1ne
other approp1iate background value and conven the in1age to a
square of size N x N where fl is an integral power of 2.

Then, all indices in the DFT 11111 fi:0111 0 to J\T - 1:

 l
F(k, l) = N
N -1 N -1
m~O
21r
·nEo f(m, n) exp - j N (mk
[ l
+ nl) ,
(2.44)

with k = 0, 1, ... , N - l, and l = 0, 1, ... , N - l.

l
f(m, n) = N
N -1 N -1
k~O ~
1 0
[ 21r
F(k, l) exp +j N (mk + nl) .
l
(2.45)

The nonnalization factor has been divided equally between the

fo1ward and inverse transfonns to be 1~, for the sake of syu1111etty.
The rectangle function and its Fourier transform:

2D function with a rectangular base of size X x Y and height A:

f(x,y) ==A if0<x<X;0<y<Y (2.46)

== 0 otherwise.
 F(u,v) ==
sin(1ruX) .
AXY [ 1ruX exp(-11ruX)
l
sin (1rvY)
x [ 1rvy exp( -j1rvY) .
l (2.47)

The Foluier transfo1m of a real iruage is. in general.

a complex function.

The exp[] fhnctions in Equation 2.47 indicate the

phase components of the spectnun.

An itnage with even syinn1et1y about the 01igi11 will have a

real Foluier transfonn.
 l if lxl < ½, IYI < ½
rect(x, y) = (2.48)
0 if lxl > ½, IYI > ½•

The Fourier transfonn of the rect function is the sine function:

rect(x, y) ¢=? sinc(u, v). (2.49)

¢:? forward and inverse Fourier-transfonn pair.

. . . sin(1ru) sin(1rv)
sinc(u,v) = sinc(u) sinc(v) = - - - - - -
1ru 1fV
(2.50)
 (a) (b)

(c) (d)

(o) (r)

Piguro 2.26: (a) Rectangle image, with tot.al size 128 x 128 pLxels and a rectangle (square) or size 40 x -to pLxcls.
(b) Log-magnitude spocmnn or Lhc inmgc in (a). (e) Rccuu1glo size 20 x 20 pixels. (d) Log-magnitude spoclrum
or the image in (c). (c) Rectangle size 10 x 10 pixels. (f) Log-magnimdc spoctmm or the image i.n (c). The
spectra have been scaled lo map the range 15. 12) to the display range (0. 255). Sec also Figures 2.28 and 2.29.
 I
(0, 0) (U/2 -.,,1,, 0) (U, 0) (0, V/2)
,,, ' ,, '- /
, ' ,, '
I
u
I

(0, V/2) (-U/2, 0) (0, 0) (U/2, , 0 )

- ' ,, - - - - - - - r - - - - - - - ... -
; ' '-
,, '
/
'
; '
' '
1 u
I

(0,
... -V/2)
I
"' , '
/

ro: vJ I (U,
;
V)
V V

(a) (b)

Figure 2.27: Frequency coordinaws in (a) the unshift.cd mode and (b} the shifted mode of display of image
spectra... i· and \' reprc:s!nt the sampling frequencies along the two axes. Spectra of images \1,ith real va.lucs
possess conjugate symmetry about U/2 and V/2. Spectra of sampled images arc periodic. with the periods equal
to U and V along the two axes. h is common practic-c to display one romplct.c period of the shifted spectrum.
including the oonjugat.c symmetric parts. as in (b). Sec also i.igurc 2.28.
 (a) (b)

l,igurc 2.28: (a) Log-magnitude spcctnuo of the rretaugle imago in Figure 2.26 (c) 9.ithout shifting. Most Fl<'f
routin~ provide spectral data it1 this format. (b) The spc.'Ctrum in (a) shifU?d or foldod such that (u. t) (0. 0)
is at, the ccmcr. It is common practice to display one complete period of the shifted spectrum. including the
conjugate symmetric pruts. as in (b). Sec also Figure 2.27.
 ,,. . ) 1(:4

,:
'"
,,.
~ Ot

(s.) (b)

1-~rure 2.29: (s.) l\lesh pfot of the rectangle image in Figure 2.26 (c). with total size L28x L28 pixels and a rectangle
(square) of size 10 x 10 pb;cls. (b) Magnitude spcetmm of the image in (a).
 I
I
I
I
111111111,
I
I
I
I

(n) (b)

(c) (d)

(c) (f)

J1igurc 2.30: (a) Rectangle image, with tot.al size 128 x 128 pixels and a rectangle of size 10 x •10 pLxels.
(b) Log-magnitude spectrum of the image in (a). (e) Rcct.a11gle siw 40 x 10 pixels; this imagl! may be oonsidcrcd
to be that in (a) rotatoo by 90". (d) Log-magnitude spectrum of the imQgc in (c). (e) Image in (e) rot,ncd by 45°
using ricarcst-neighbor sclcction. (f) Log-magnitude spectrum of the imago in (e). Spectra scaled to map [5, 121
to the displa\" rru1J?.c IO. 2551.
The circle function and its Fourier transform:

Circular apertures and fimctions are encountered often in itnaging

and ilnage processing.

. (r )
circ = l 1 if r < 1 (2.51)
0 if r > 1,

where r = J(x2 + y2).

i
The Fotuier transfonn of circ(r) is J1(21rv).

where v = J(u 2 + v2) represents radial frequency in the 2D

(u, v) plane. and 11 is the first-order Bessel function of the
first kind.
 (11) (b)

Figure 2.3 I: (a) Image of a circulAr disc. The radius of the disc is IO pLxck the size of the image is 128 x 128
pixels. (b) Log-magnitude spectrum of tho image in (a). Seo also Figur~ 2.32 and 2.33.
 (a) (b)

Figure 2.32: (a) ~lcsh plot. of the circular disc in Figure 2.3l (a). The radius of the disc is JO pixels: the size or
the image is 128 x J~ pLxcls. (b) ~1agnitude 3Pcctrum or the image in (a).
 11 11

tl 1C

--·
IQ

(a)
IQ •OD 1:10 10

--
l0

(b)
C)

Figure 2.33: (a) Profile of the log-magnitude ~cctrum in ~gure 2.31 (b) Along the central horizontal a.xis.
(b) Profile in (a) shown only for positi\·e frequencies. The frequency a.xis is indicated in samples: the true
frequency values depend upon the sampling frequency.
Profiles of 2D syste1n transfer functions:

it is connnon to show only one half of the profile for

positive frequencies.

It is to be assumed that the system possesses axial or rotational

synnnetiy: that is. the system is ;sonvpic.
 (11) (b)

(c) (d)

Figure 2.34: (a) TEM image of collagen fibers in 11. nonnnl rabbit ligament sample. (b) Log-magnitude spectrum
of the image in (a). (c) T&\I image of collagen fibers in a scar tissue sample. (d) Log-magnitude spectrum of the
image ir1 (c). Sec also Figure 1.5 and Section 1.4.
 (a) (b)

(c) (d)

l<'iguro 2.35: (a) SEM image of collagen fibers in a norrnlll rabbit ligament sample. (b) Log-magnitude spocmnn
of the image in (a). (c) SEM imnge of collagen fibers in a scar tissue sample. {d) Log-mB!!Ilitude spocmun of the
image in (c). Sec also Figure 1.8 and Section 1.4.
Important properties of the Fourier transform (FT)

1. The ke1uel of the FT is separable and synunetric.

Facilitates the evaluation of the 2D DFT as a set of ID row
transfonns. followed by a set of ID column transfonns.

1
F(k,l) == N x (2.52)

N-1 ( 21r ' N-1 ( 21r '

E exp -j~mkl E f(m,n) exp -jNnll.
m=O \ jv ) n=O \ )
1D FFT routines 1nay be used to obtain 2D/ Nill FT:

_ r r~ N-1 (- .21r ,1
F(m, l) - 1\ IN L f(m, n) exp J I\Tnl 11, (2.53)
l n=O j )J

1 N-1 ( 21r )
F(k,l) == NmLo F(m,l)exp - j Nmk . (2.54)

Check if 1{, is included in the fo1ward or inverse 1D FFT.

2. Parseval's theorem:
The Fourier transfonn is an energy-conse1ving transfonn.

IJ(x, Y)l 2 dx dy
00
fx~-oc /y -oo

= fu~-oo 00
fv -oc IF(u, v)l 2 dudv. (2.55)

3. The inverse Fourier transfonn (IFT) rnay be pe1fonned using

the sarne FFT routine by taking the fo1ward Fourier transfonn
of the complex conjugate of the given function. and then
taking the co1nplex conjugate of the result:

f == I FT(F) == [FT{F*}]*
4. The Fourier transfo1m is a linear transfonn.
Images are often con11pted by additive noise:

g(x, y) = f(x, y) + TJ(x, y). (2.56)

G(u, v) = F(u, v) + rJ(u, v). (2.57)

1V1ost real-life images have a large po1tion of their energy

concentrated at (u, v) = (0, 0) in a low-frequency region.
Edges, sha1p feanires, and small-scale or fine details lead to
increased strength of high-frequency components.
Random noise has a spectnuu that is equally spread all over
the frequency space (flat, unifonn. or "white" spectnllll).
Indisc1iminate removal of high-frequency components could
cause bluning of edges and the loss of the fine details.
5. The DFT and its inverse are pe1iodic signals:

F(k, l) == F(k ± aN, l) == F(k, l ± aN)

== F(k ± af\l, l ± {3N), (2.58)

where a and f3 are integers.

6. The Fotuier transfonn is conjugate-symrnetiic for i1nages with
real values:

F(-k, -l) = F*(k, l). (2.59)

IF(-k, -l)I = IF(k, l)I,

LF(-k, -l) = -LF(k, l).

The 1nagnitude spectnun is even syminet1ic.

The phase spectnun is odd syinmet:ric.
7. A spatial shift or translation applied to an image leads to an
additional linear phase cornponent in its Fourier transfonn.
The 1nagnitude spectnllll is unaffected.
If f(m, n) ¢=? F(k, l) are a Fornier-transfonnpair. we have

2
f (m - mo, n - no) <=} F(k, l) exp f-j N7r (kmo + lno)1,
l J
(2.60)

where (m 0, n 0) is the shift applied in the space dornain.

Conversely, we also have

f .21r l
J(m, n) exp IJ N (k 0 m + l 0 n)I ¢:? F(k - k0 , l - l 0 ).
l J
(2.61)

This prope1ty has ilnpo1tant implications in the 1nodulation of

1D signals for transmission and collllnunication:
it does not have a sintilar application with 2D ilnages.
8. F(O, 0) gives the average value of the iinage:
a scale factor 1nay be requi.J.·ed depending upon the definition
of the DFTused.
9. For display ptuposes. log 10 [1 + IF(k, l)l 2] is often used:
the addition of unity (to avoid taking the log of zero), and the
squaiing 1nay s01neti.J.nes be dropped.
It is also conunon to fold or shift the spect11un to bring the
(0, 0) frequency point (the "DC" poiI1t) to the center.
Folding of the spectnun could be achieved by multiplying the
iinage f(m, n) with (-l)(m+n) before the FFT is co1nputed.
Because the iI1dices m and n are integers. this a1notu1ts to
merely changing the signs of alten1ate pixels.
This is related to the property in Equation 2. 61 ~ritb
ko == lo == N /2:

exp [j; (k m + l n)]

0 0 = exp[j1r(m + n)] = (-l)(m+n),
(2.62)

f(m,n) (-l)(m+n) ¢:;> F(k - N/2,l- N/2). (2.63)

10. Rotation of an iinage leads to a co1Tesponding rotation of the
Foluier spect:11un.

f ( m 1, n 1) ¢:;> F (k 1, l1) , ( 2. G4)

m1 == m cos 0 + nsin 0; n1 == -msin 0 + n cos 0;

(2.65)

k1 == k cos 0 + l sin 0; l1 == -k sin 0 + l cos 0. (2.66)

11. Scaling an in1age leads to an inverse scaling of its
Foluier transfo1m:

1 (k l)
J(am,bn) {:} labl F a'b' (2.67)

where a and bare scalar scaling factors.

The shrinking of an i1nage leads to an expansion of its
spectnun. with increased high-fi:equency content.
On the contra1y, if an irnage is enlarged, its spectnun is
shn1u.k:, with reduced high-frequency energy.
12. Linear shift-invariant syste1ns and convolution:
Most i1naging systenlS may be modeled as linear and
shift-invariant or position-invruiant systems that are
completely characterized by their PSF s.
The output of such a system is given as the convolution of the
input itnage with the PSF:

g(m, n) == h(m, n) * f(m, n) (2.G8)

N-1 N-1
== E E h(a, {3) f(m - a, n - {3).
o=O {3=0
Upon Fou1ier transfonnation, the convolution rnaps to the
1nultiplication of the two spectra:

G(k,l) ~ H(k,l)F(k,l). (2.69)

h(x, y) * f(x, y) ¢:;> H(u, v) F(u, v), (2.70)

expressed now in the continuous coordinates

(x, y) and (u, v).
The convolution ¢:;> rnultiplication property with the DFT
ilnplies periodic or circular convolution:
Circular convolution 1nay be 1nade to be equivalent to linear
convolution by zero-padding.
13. Multiplication of ilnages in the space do1nai11 is equivalent to
the convolution of their Fotu·ier transfonns:

(2. 71)

So1ne types of noise get nntltiplied with the i1nage.

When a transparency, such as an X-ray image on fihn. is
vie~red usi11g a light box. the resttlting image g( x, y) n1ay be
1nodeled as the product of the transparency or transnlinance
ftu1ction f (x, y) with the light source intensity field s(x, y):
g(x, y) == J(x, y) s(x, y).
Ifs( x, y) is absolutely unifo1111 with a value A, its Fourier
transf011n will be an impulse:
S(u, v) == A 8(u, v).
The convolution of F ( u, v) with A 8 ( u, v) will have no
effect on the spect1111n except scaling by the constant A.
If the sotu·ce is not unifo1111. tl1e viewed image
._ will be a

disto1ted version of the 01iginal:

G(u, v) == F(u, v) * S(u, v).
14. The con·elation oftv.10 images f(m, n) and g(m, n) is

1V-l JV-1
iJ.g(a, 6) == E
1 E f(m, n) g(m+a, n+/3). (2.72)
m=O n=O

Con·elation is useful in the comparison of nnages where

feanu·es that are collllnon to the images rnay be present with a
spatial shift (a, /3).

r J.g(k, l) == F(k, l) G*(k, l). (2.73)

A related 1neasure, la1own as the correlaholl coefficient and
useful in te,nplate ,natching and image classification, is

E"'IV=l
m-0 EJV-l
n-0 f(m. n) g(m. n)
i ,

' ==- • J'>(

JV-1 N-1
[Em=O En=O - m, n ) N-1 En=O
Em=O N-1 g~'>( m, n )]2
(2.74)

Here. it is assun1ed that the two itnages f and g are aligned

and registered. and are of the sa1ne scale and orientation.
15. Differentiation of an i1nage results in the extraction of edges
and highpass filtering:

f)J~';y)? j21ru F(u,v);

8f(x,y) .
By <=> J21rv F(u, v). (2.75)

Gain of the filter increases linearly with frequency u or v.

Delivatives are approxilnated by differences:

Jy(m, n) ~ f(m, n) - f(m - 1, n),

fx(m, n) ~ f(m, n) - f(m, n - l), (2.76)

(using matrix notation).

Operators based upon differences could cause negative pixel
values in the result.
In order to display the result as an iinage, it will be necessa1y
to map the full range of the pixel values, including the
negative values, to the display range available.
The 1nagnitude of the result may also be displayed if the sign
of the result is not important.
Differentiation results in the re1noval of the intensity
infonnation from the ilnage.
The values of the spectn1m for u == 0 or v == 0 are set to zero.
 (a) (b)

(c) (d)

I
I
I

-= I
I
I
I
I
I

(c) '
(r)

i;'igurc 2.36: (a) Image or a roctangular box. (c) Horizontal and (e) vertical dcrirntim, or the image in (a).
respccti\'cly. (b). (d), and (r): Log-magnitude spectra or the images in (a). (c). and (e). respectively. The images
in (c) and (e) were obtained by mapping the range l-200, 2001 1-0 the display range or (0, 2551. Negative differences
appc,-u in black. positive differences in white. The "l)CCtra show values in the range (5, 121 mapped to I0, 255).
 (a) (b)

(c) (d)

(c) (f)

Figure 2.37: (a) Image of a myocyui. (c) Horizontal and (o) \Utical derivatives of tho image in (a). rcspcctivcly.
(b). (d). and (f): Log-magnitude spcctm of the images in (a), (c), and (c), rCl>-pocti\-cly. Images in (c) and (c)
were obtained by mapping tho rnngo l-20, 20J LO tho di'iJ)lay range of [O. 255J. The spcctrn show \'lllucs in the
range (3. 12) mapped to (0. 255).
 ~6 ••, .- •. \
L_·,2_ ._.·.
-_-·
6

1'.. .·• ;,,:·

• ·.. ·•

,':~\,iii '

(c) (d)

(e) (r)

Figure 2.38: (a) MR image or n lmcc. (c) llorizontal and (c) vcnical dcriv-aLivcs or the image in (a). rcspoctivcly.
{b). (d). and (f): Log-magnitude spcctrn of the inLQgCS in (a). (e). 9.lld (c}. rcspcctivdy. The images in (e) and
(e) were obtained by mapping the rruil,rc [-50, 50) to the display range of [O. 255). Negative differences appear in
black, pa;itivc differences in white. ·nic spectra show values in the range (3. 12] mapped 10 [O, 255[.
16. The Laplac;an of an in1age:

? a2J a2J
v-J(x,y) ==a? +-a9· (2.77)
x- y-

(2.78)

The spectnun of the ilnage is 1nultiplied by (u 2 + v 2), which

is isotropic and i11creases quadratically with frequency.
High-:frequency components are a1nplified.
Otnnidirectional operator: detects edges in all di.J.·ections.
The second de1ivatives 1nay be approxi1nated as follows:
Talcing the delivative of the expression for f~( m, n) in
Equation 2.76 for the second tin1e. we get

II

Jy(m, n)

~ f (m, n) - f (m - 1, n) - [f (m - 1, n) - f (m - 2, n)]

= f(m, n) - 2 f(m - 1, n) + f(m - 2, n) (2.79)

(using matrix notation).

Causality is usually not of concen1 in ilnage processing:
desil·able to have operators use collections of pixels that are
centered about the pixel being processed.
Applying a shift of one pixel to the result above, we get
If

Jy(m, n) (2.80)

~ f(m+ 1,n) -2f(m,n) + f(m-1,n)

== f(m - 1, n) - 2 f(m, n) + f(m + 1, n).

If

fx(m, n) ~ J(m, n - 1) - 2 J(m, n) + f(m, n + l).

(2.81)
The Laplacian could then be ilnple1nented as

!L(m, n) == f(m - 1, n) + f(m, n - 1) - 4f(m, n)

+ f(m + 1, n) + f(m, n + l). (2.82)

convolving the ilnage with the 3 x 3 1nask or operator

0 1 0
1 -4 1 (2.83)
lO 1 0J
 Ill

■
111
,,,

1~111111111
.
I
111111i111111
'
,,,I
,,,
Ill
(a) (b)

J<'iguro 2.39: (a) Laplacian of the roctanglo image i11 Figure 2.36 (a). (b) Log-magniLude SJX.'CLrum of the image
in (a).

(a) (b)

Figuro 2.-IO: (a) Laplacian of the myocyte image in Figure 2.37 (a). (b) Log-magniLude spectrum of tho image in
(a.).
 (a) (b)

l•igurc 2.41: (a) Laplacian of the .\1R image in Figure 2.38 (a). (b) Log-magnitude spectrum of the image in (a).
17. Integration of an irnage leads to s1noothing or bluning, and
lowpass filtering:

1
J:: -oo J(a, y) da. {cc} . F(u, v), (2.84)
J 21fU

1
/% -oo f (x, (3) d(J {cc} . F(u, v). (2.85)
J27fV

The weighting factors that apply to F (u, v) dintinish with

increasing frequency. and hence high-frequency co1nponents
are attenuated by this operation.
The integration of an iniage fro1n -oo to the cu1Tent x or y
position is seldo1n encountered in practice.
Instead, it is co1runon to encounter the integration of an ilnage
over a small region or apern1re su1Toundit1g the ClUTent
position, in the f01m

_ 1 A/2 B/2
g(x, Y) - AB fc~=-A/2 !(3=-B/2 f (x + a, y + (3) da d{3,
(2.86)

where the region of integration is a rectangle of size A x B.

The no1n1alization factor A~ leads to the average intensity
being computed over the area of integration.
This 1nay be interpreted as a 111oving-average (MA) filter.
Averaging over a 3 x 3 aperture or neighborhood:

1 1 1
g(m, n) == - E E f(m + a, n + {3). (2.87)
9 a=-1 {3=-1

1
g(m, n) == x
9

[ f(m-l,n-l) +f(m-l,n) +f(m-l,n+l)

+f(m,n-l) +f(m,n) +f(m,n+l)
+f(m+l,n-l) +f(m+l,n) +f(m+l,n+l) ].
(2.88)
Equivalent to convolution of the irnage f(m, n) with

1 1 1
1
- 1 1 1 (2.89)
9
1 1 1

which 1nay be viewed as the PSF of a filter.

Equivalent to rnultiplication of the Fourier transfonn of the
irnage with a 2D siric function.
Integration or averaging only along the horizontal or ve1tical
directions 1nay be perfonned via convolution with the aiTays
½[1 , 1, 1] or ½[1, 1, 1] T.
 (a) (b)

I
I
I
,1
1illl~1
,,
I
I
I

(c) (d)

(c) (f)

Figure 2.-12: (a) lmngc of a roctanguh1r box. Raiults of averaging using throo pixels in the (c) horizomru and
(c) vcnical directions. rcspcctivcly. (b), (d), and (f): Log-magnitude !>l)Cctrn of the imagl'S in (a). (c). and (e).
rcspocti\'l!ly. The spectra show \'ruuos in the rruJgc i5, 121 mapped 10 (0. 255.
 -- - - -

(a) (b)

F'i.!,ruro 2..13: (a) Result, of 3 x 3 averaging of the rectangle image in Figure 2.42 (a). (b) Log-mAgDiLude spoctrwn
of the image in (a).
 (a) (b)

(c) (d)

(c) (f)

Piguro 2.44: (a) Image of a myocytc. Results of averaging using throo pixels in the (c) horizomal and (c) rnrtical
direction,. rcspoctivcly. (b), (d), and (f): Log-nmgnirndc Sp(.'Clm of the imagos in (a), (c). and (o). respectively.
The spcctrn show vruuos in the rnngc 13. 121 mapped to 10. 255j.
 (a) (b)

l<'igurc 2A5: (a.) Rc.5ult of 3 x 3 averaging of lhc myocyte imRgc in Figure 2.44 (a.). (b) Log-magnitud~ spectrum
of the imll'JC in (a).
 (a) (b)

(c) (d)

(c) (f)

Figure 2.46: (a) ~Ill image of a knee. Rosultsof B\'ernging using throe pixels in the (c) horizont;il and (c) \'t'rtical
directions. rcspocLin:ly. (b), (d). and (f): log-rnagnitudc spectra of the images in (a), (c). illld (c). rcspoctivcly.
1110 !i'J)CCtrs show values in the range (3. J2j nuippcd to (0. 255j.
 .

(a)
• -

(b)
.

(11) Result of 3 x 3 avcrAging of the knee ~tR imAgc in r,'igurc '2.'16 (a). (b) Log-lThlgnitudc spectrum
J.igurc 2..17:
of the image in (a).