Unit 3

Biomedical Image denoising

Introduction
• Biomedical images are often affected and corrupted by various types of
noise and artifact

• Any image, pattern or signal other than that of interest could be termed as
interference, artifact or simply noise

• The sources of noise could be physiological, the instrumentation used, or

the environment of the experiment

• The problems caused by artifacts in biomedical images are vast in scope

and variety
• their potential for degrading the performance of the most sophisticated
image processing algorithms is high

• The removal of artifacts without causing any distortion or loss of the

desired information in the image of interest is often a significant challenge
Characterization of Artifacts
Random noise
• The term random noise refers to an interference that arises
from a random process such as thermal noise in electronic
devices and the counting of photons
• A random process is characterized by the PDF representing
the probabilities of occurrence of all possible values of a
random variable
Co11Sidcr a randoru process '1 that is characteriied by t.he PDF p,1 (11). The
vrocess could be a function of tiwc as 11( t), or of svacc iu lD, 2D, or 3D
as 11 (:e), r,( :e, y), or r,( :e, y, .:) ; it could also be a spatio-tcruvoral function as
11( :e, y, .:, t). The arguwcnt of the PDF represents the value that the rand ow
process can assuwc, which could be a voltage in the case of a function of tiwc,
or a gray level in the case of a 2D or 3D iwage. The use of t.hc sawc sywbol
for the function aud the value it can assuwc when dealing with PDFs is useti1l
when dealing wit.h several randow processes.
 The n1can 11,1 of thr ra11dou1 process 11 is given by the Jirst-order n10111ent
of the PDF, defined as

(3.1)

where E[η] represents the statistical expectation operator

It is common to assume the mean of a random noise process to be zero

The mean-squared (MS) value of the random process η is given by the second-
order moment of the PDF, defined as

(3.2)
The variance a~ of the process is de.fined as the second central ruoruent.:

(3.3)

The square root of the variauce gives the standard deviatiou (SD) a,1 of the
process. I\ ote that a;, = 1!J _11 2 - µ;,. lf thr wean is ~ero, it follows that
a~ H.11 2 , that is, the variance and the ~1S values arc the sanie.
 I I I
0.25 ... -

0.2,..

0.15

0.1

0.05 ~
I o
I

i --0.ai ... I
--0.1 ...

-0.15 .,_

--0.2 ...

--0.25 ...
I I I I I

50 100 150 200 250

sample numller

FIGURE 3.1
A time series coruposed of random noise samples with a Gaussian PDF having
µ - 0 and a 2 - 0.01. ~tS value - 0.01; HJvlS - 0.1. See also Figw·es 3.2 and
3.3.
• An example of a random function of time is the current generated
by a CCD detector element due to thermal noise when no light is
falling on the detector (known as the dark current)

• The statistical measures described above then have physical

meaning: the mean represents the DC component, the MS value
represents the average power, and the square root of the mean-
squared value (the root mean squared or RMS value) gives the
average noise magnitude

• These measures are useful in calculating the SNR, which is

commonly defined as the ratio of the peak-to-peak amplitude range
of the signal to the RMS value of the noise, or as the ratio of the
average power of the desired signal to that of the noise

• Special-purpose CCD detectors are cooled by circulating cold air,

water, or liquid nitrogen to reduce thermal noise and improve the
SNR
 • When the values of a random process η form a 2D function of space, we have a noise
image η(x,y)

FIGURE 3.2
An image cow posed of random noise samples with a Gaussian l' UF having
µ - 0 and a 2 - 0.01. ~IS value = 0.01; B~IS 0.1. The normalized
pixel values in the range . 0.5, 0.5_ were linearly mapped to the display range
_0, 255_. See also Figm·e 3.3.
0.012

I o.o,
I
~ 0.008
l!'
2

hooo

0.004

0.002

FIGURE 3.3
L - = = =
--- o ~ ~ ~ ~ ~

.\lormalized histogram of lhc image in Figure 3.2. The samples were generated
using a Gaussian process withµ= 0 and a 2 = 0.01. ~IS value= 0.01; HMS
- 0.1. Sec a Lso Fi gurcs 3.1 and 3. 2.
• Several possibilities arise in this situation

• We may have a single random process that generates random gray

levels that are then placed at various locations in the (x, y) plane in
some structured or random sequence

• We may have an array of detectors with one detector per pixel of a

digital image, the gray level generated by each detector may then
be viewed as a distinct random process that is independent of
those of the other detectors

• A TV image generated by such a camera in the presence of no input

image could be considered to be a noise process in (x,y,t) that is a
function of space and time
• A biomedical image of interest f(x, y) may also, for the sake of generality be
considered to be a realization of a random process f

• Such a representation allows for the statistical characterization of sample-to-

sample or person-to-person variations in a collection of images of the same organ,
system, or type

• For example, although almost all CT images of the brain show the familiar cerebral
structure, variations do exist from one person to another
• A brain CT image may be represented as a random process that exhibits certain
characteristics on the average
• Statistical averages representing populations of images of a certain type are useful
in designing filters data compression techniques, and pattern classification
procedures that are optimal for the specific type of images
• However, it should be borne in mind that, in diagnostic applications, it is the
deviation from the normal or the average that is present in the image on hand that
is of critical importance
 A

\.V hen au iwage / ( x, y) is observed iu the presence of randoru noise '1, the
detected iruage g( x, y) ruay be treated as a realiz;at.ion of another randon1
process g. ln most cases, the uoise is additive, and the observed iwage is
expressed as
g(:e, y) = /(:e, y) - 17(:e, y). (3.4)
Each of the randoru processes f, 17, aud g is cbaractcriz;ed by its own PDl1..
Pt (f), p,, (17), and p" (g), respectively.

In most practical applications, the random processes representing an image of

interest and the noise affecting the image may be assumed to be statistically
independent processes

Two random processes f and η are said to be statistically independent if their

joint PDF p f, η (f, η ) is equal to the product of their individual PDFs given as
pf (f) p η(η)

It then follows that the first-order moment and second-order central moment of
the processes in Equation 3.4 are related as
µ ,, -:_ µf - }!; .1• . l (3. 5)
H(
_g (3. 6)
where µ represents the mean and a 2 represents the variance of the random
process indicated by the subscript, and it is assumed that µ,1 - 0.
 Ensemble averages: W'hen the PIJ Fs of the random processes of con-
cern are not known, it is common to approximate the statistical expectation
operation by averages computed using a collection or ensemble of sample
observations of the random process. Such averages are known as ensemble a·u-
erages. Suppose we have Mobservations of the random process / as functions
of (:e, y): /1(x, y), / 2(x, y), ... 1 fM (:e, y); see Figure 3.4. We may estimate the
mean of the process at a particular spatial location (:e 1, yi) as
rM (x, yJ

(3. 7)
rk (x, yJ

-------------+
Jal average
µ {le)
r

l':Z (X, y)

r (X, y)
1

ensemble average
µ (X , y }
r l l
The aut,ocorrelation function (AUF) q,1 (~ 1 , ~ 1 - c.1 1 Y1i y1 - ,B) of the randoru
process f is defined as

which may be cstiruatcd as

A; 1

where α and β are spatial shift parameters

If the image f(x, y) is complex, one of the versions of f(x, y) in the products
above should be conjugated; most biomedical images that are encountered
in practice are real-valued functions, and this distinction is often ignored

The ACF indicates how the values of an image at a particular spatial

location are statistically related to the values of the same image at another
shifted location
• If the process is stationary, the ACF depends only upon the
shift parameters, and may be expressed as φf(α,β)
The thi·ee equations above way be applied to siguals that are fuuctious
of tiwe by replacing the spatial variables (:e, y) with the tewporal variable
t, replacing the shift pararueter a wi t.h r to represent teruporal delay, and
waking a few other related changes.

\.Vhen µt(:e1, Y1) is coruput,cd for every spat,ial location or pixel, we get. au
average iruage that could be e..~pressed as /(:e, y). The iruage f ruay be used
t.o represent. the random process f as a prot.ot.ype. For pract.i cal use, such au
average should be coruputed using saruple observations that are of the sarue
si;t;C, scale, orientation, ct.c. Siwilarly, t.hc AUF way also be coruput.cd for all
possible values of its indices to obtain au iruage.
 Temporal and spatial averages: \.Vhen ""e have a saruple observation of
a random process /1,; ( t) as a function of time, it is possible to compute liuie
a.uei'u.ges or leuipora.L sla.lislics by i utegratiug along the ti rue axis . 31.:

l-'t(k) - liru
T-HXJ
?
T
1T/'2
-T/2
/1,;(t) dt. (3.10)

The integral would be replaced by a suuillJ.atiou in the case of sampled 01·

discrete-time signals. The time-averaged ACF </>1(r, k) is given by

l 1T,'2
</> t ( r, k) =- . lim -;-- / k ( t) f k ( t - T) dt. (3.11)
T-'!>00 T -Ti2

Similarly, given au obse1·vation of a 1·audoru process as au truage /1,,;(x, y),

we 1uay compute averages by integ1·ating over the spatial don1ain, to obtain
spatial averages or spatial statistics; see FigU1·e 3.4. The spatial rueau of the
image /k(-.e,y) is given by

l
µt(k) - A. 1 loo
00
-oo -oo /1,,;(-.e, y) d-.e dy, (3.12)

whe1·e A. is a norruali:Gatiou facto1·, such as the actual area of the iruage. Ob-
serve that the spatial ruean above is a single-valued entity (a scalai·). For a
stationary process, the spatial ACF is given by

(3.13)
Asuitable norn1aliiation factor, such as the total energy of the in1age _which is
equal to ~1(0, O)_ UJay br included, if necessary. The san1plc index k bcconies
irrelevant if only one observation is available. ln practice, the integrals change
to sununat.ions over t.hc spare of the digital iu1agc available.
\1Vhcn ""e havr a 2D image as a funct.ion of time, such as IV, video, tluo-
roscopy, and cine-angiography signals, we have a spatio-temporal signal that
niay be expressed as f (:e, y, t ); sec Figure 3.5. \Ve n1ay then co111putc statis-
tics over a single frame /(-.e, y,t 1 ) at the instant of tinJc t1 , which arc known
as intrafran1c statistics. \,Ve could also compute paraUJetcrs through 111uH.iplc
fran1es over a certain period of tin1c, which arc called interfran1c statistics;
thr signal over a specific period of tin1e niay then be treated as a 3D dataset.

t (x, y, c)

Random functions of time may thus be

characterized in terms of ensemble and/or
temporal statistics t (x, y, C )
l

-------------• spacial or
intra frame

Random functions of space may be scaciscics

represented by their ensemble and/or spatial

statistics

cenporal or incertrame scaciscics

• The mean of an image represents its average intensity or
density; removal of the mean leads to an image with only the
edges and the fluctuations about the mean being depicted

• The ACF plays an important role in the characterization of

random processes

• The Fourier transform of the ACF is the power spectral density

(PSD) function which is useful in frequency-domain analysis

• Statistical functions as above are useful in the analysis of the

behavior of random processes and in modeling spectrum
analysis, filter design, data compression, and data
communication
Examples of noise PDFs
• Gaussian noise PDF
• The most commonly encountered and used noise PDF is the Gaussian or normal
PDF, expressed as

exp [ (3.14)

• ?
A Gaussian PDF is cowplet.ely specified by its wean µa: and vanance a:.
Figure 3.6 sho,..vs three Gaussian PDFs with 1.1. - 0, a - I; 1.1. - 0, a - 2; and
µ - 3, a - 1. See also Figw:es 3.2 and 3.3.
\.Vhen we have two jointly norrual randoru processes ~ and y, the biva1:iate
norrual PDF is given by
1
PJ:.y(~,y) X
, / 41r 2 ( l ...,-2) Ua;Uy

exp {
l
2(1 - ...,- 2 )
[ (x "t)2
2...,-(x µJ;)(y
U-a;Uy
µy) (y
"?)2 J}' (3.15)

where ,..,. is the cor.relat.ion coellicient. given by

(3.16)

lf ,..,. - 0, the t.wo p1·ocesses a1·e uncon.-elat.ed. The bivadat.e no1·rual PDF then
reduces to a product of two univariate Gaussians, which implies that the two
processes ai·e statistically independent..
 0.35

0.3

0.25
Q)
~
0
n.
;
111
(I)
:::,
0.2
I
/
-- ' \
(U I \
0
0.15 I \
I \
I \
I
0.1

. \
0.05
I
I .. '\
/ '
' ....

-8 -6 -2 0 2 d 6 8 10
)(

FIGURE 3.6
Three Gaussian P D.Fs. Solid line: µ - 0, a - l. Dashed line: µ - 0, a - 2.
Dotted line: µ - 3, a - 1.
 Uniform: All possible values of a uuiforwly distributed raudoru process
have equal probability of occurrence. The PDF of such a randolli process over
the rauge ( u, b) is a rectaugle of height Cb~ u) over the rauge ( u, b). The rue au of
the process is (u;b), aud the variance is (b~~):!. Figtu-e 3.7 shows two llllifo1·U1
PD Fs correspoudiug to raudoru processes with values spread over the raugcs
( 10, 10) and ( - 5, 5). The quanti:Gatiou of gray levels in an iwage to a fiwte
uuruber of integers leads to an error or noise that is uuiforllily distributed.

0.12

0.1 ----- - ------ -- - ,

0.08

..
u.
C
0..
E o.06
0
=C:
:::,

0.04

0.02

I
I
I
0
-10 -S 0 s 10
X

FIGURE 3.7
Two uniform PDFs. Solid line: µ - 0, range - ( 10, 10). Dashed line: 11 - 0,
range - ( 5, 5).
Poisson noise PDF
• The counting of discrete random events such as the number
of photons emitted by a source or detected by a sensor in a
given interval of time leads to a random variable with a
Poisson PDF

• The discrete nature of the packets of energy (that is, photons)

and the statistical randomness in their emission and detection
contribute to uncertainty, which is reflected as quantum
noise, photon noise, mottle, or Poisson noise in images

• Shot noise in electronic devices may also be modeled as

Poisson noise
 ()ne of the forruulatious of the Poissou PDF is as follo"vs: The vrobability
that k photous arc detected iu a ce1taiu iutc1·val is given by

(3.17)

Herc, µ is the mean of the process, which represents the average nuruber of
vhot.ous couutcd iu the svccificd inter val over ruany trials. The values of JJ ( k)
for all (iutcgcr) k is the Poissou PDF. The variance of the Poissou PDF is
equal to its u1cau.
 The Poisson P Dl'1., tends toward the Uaussiau PDF fo1· lai·ge wean values.
Figure 3.8 shows two Poisson PDFs aloug with the c:aussiaus for the sawc
paran1etcrs; it is seen that the Poisson aud Uaussiau PDFs for µ - a 2 - 20
ruatch each other well.

0.2
I \
0.18 I
1

0.16 I

3!.
~ 0.14
C
"'
~ 0.12
"'
C)

.§
~ 0.1

I o.oo
0.06

0.0I

0.0'2

0
0 5 10 15 20 25 30 40
k

FIGURE 3.8
Two Poisson PDFs with the corresponding Gaussian PDFs superimposed.
2
Bars with x and., dashed envelope: µ - a - 4. Bars with o and solid
envelope: µ = a- = 20.
Laplacian Noise PDF
The Laplacian PDF is given by the function

Px(x) - l
2 0-x
exp { (:1.18)

where µx. and er~ are the rnean and variance, respectively, of the process.

0.7

0.6

0.5

(/)

~ 0.4
ti I
al I I
:§- 0.3 I I
I I
I

0.2

0.1

--8 --6 -4 -2 0 2 4 6 8 10
X

FIGURE 3.9
Two Laplacian PDFs withµ,= 0, a 2 = 1 (solid) andµ,= 0, a 2 = 4 ( dashed)
Rayleigh Noise PDF
The Rayleigh PDF is given by the function
2
2 (x a) }
Px (x) - b (x b
u(x a)' (~ .19)

where ·u(x) is the unit step function such that

lifx>O
U X
( )
- -
{ 0 otherwise. (3 .20)

The mean and variance of the Rayleigh PDF are determined by the pararneters
a and b as l173] µx - a+ ✓(1rb/4) and a~ - b( 4 1r)/4.
 Figu1·e 3.10 shows a H.ayleigb PDF ""itb u - l and b - 4. The H.ayleigb
PDF bas been used to ruodel speckle noise _175_.

0.4

0.35

0.3

u..
~ 0.25
&
-«i
i 0.2

0.15

0.05

00 1 2 3 4 5 6 7 8
X

FIGURE 3.10
H.ayleigb PDF with a - 1 and b - 4.
Structured noise
Power-line interference at 50 Hz or 60 Hz is a common artifact in many
biomedical signals

The typical waveform of the interference is known in advance and, hence, it

could be considered to be a form of structured noise

It should, however, be noted that the phase of the interfering waveform will
not usually be known

Furthermore, the interfering waveform may not be an exact sinusoid; this is

indicated by the presence of harmonics of the fundamental 50 Hz or 60 Hz
component

Notch and comb filters may be used to remove power-line artifact

A common form of structured noise in biomedical images is the grid
artifact

On rare occasions, the grid used in X-ray imaging may not translate or
oscillate as designed; the stationary grid then casts its shadow, which is
superimposed upon the image of the subject

Stationary grids with parallel strips create an artifact that is a set of

parallel lines which is also periodic

Depending upon the strip width, strip orientation, and sampling

resolution, the artifactual lines may appear as thin straight lines, or as
relatively thick rectangular stripes with some internal grayscale variation
In some imaging applications, grids and frames may be used to position
the object or subject being imaged

The image of the grid or frame could serve a useful purpose in image
registration and calibration, as well as in the modeling and removal of
geometric distortion

The same patterns may be considered to be artifacts in other

applications of image processing

Labels indicating patient identification, patient positioning, imaging

parameters, and other details are routinely used in X-ray imaging
Ideally, the image cast by such labels should be well removed from the
image of the patient on the acquired image, although, occasionally they
may get connected or even overlap

Such items interfere in image analysis when the procedure is applied to

the entire image
Preprocessing techniques are then required to recognize and remove
such artifacts

Surgical implants such as staples, pins, and screws create difficulties and
artifacts in X-ray, MR, CT, and ultrasound imaging

The advantage with such artifacts is that the precise composition and
geometry of the implants are known by design and the manufacturers
specifications

Methods may then be designed to remove each specific artifact

Physiological interference

the human body is a complex conglomeration of several systems and processes

Several physiological processes could be active at a given instant of time, each one
affecting the system or process of interest in diverse ways

A patient or experimental subject may not be able to exercise control on all of his
or her physiological processes and systems

The effect of systems or processes other than those of interest on the image being
acquired may be termed as physiological interference
several examples are listed below

• Effect of breathing ou a chest X-ray iruage.

• Effect of breathing, peristalsis, aud woveweut of ruaterial through the

gastro-iutcstiual systeru ou CT iruages of the abdorucu.

• Effect of cardiovascular activity ou CT iruages of the chest.

• Effect of vulsatile ruoveweut of arteries iu subtraction angiography.

Physiological interference may not be characterized by any specific waveform,
pattern, or spectral content, and is typically dynamic and non-stationary

Thus, simple, linear band-pass filters will usually not be effective in removing
physiological interference

Normal anatomical details such as the ribs in chest X-ray images and the skull
in brain imaging may also be considered to be artifacts when other details in
such images are of primary interest

Methods may need to be developed to remove their effects before the details
of interest may be analyzed
Other types of noise and artifact
Systematic errors are caused by several factors such as geometric distortion, mis-
calibration, nonlinear response of detectors, sampling, and quantization

Such errors may be modeled from a knowledge of the corresponding parameters,

which may be determined from specifications, measured experimentally or derived
mathematically

few other types of artifact that cannot be easily categorized into the groups discussed
above are the following
• Punct.at.e or shot. noise due t.o dust. on the screen, film, or examination
table.

• Scratches on filru that. could appear as intense line segruent.s.

• Shot. noise due t.o inactive eleruent.s in a detector array.

• Salt-and-pepper noise due t.o impulsive noise, leading t.o black or white
pLx.els at. t.he extreme ends oft.he pLx.el-value range.

• Film-grain noise due t.o scanning of films wit.h high resolution.

• Punct.at.e noise in chest. X-ray or ruammographic images caused by cos-

ruet.ic powder or deodorant. ( which could masquerade as ruicrocalcifica-
tions).

• Superiruposed images of clothing accessories such as pins, hooks, but.-

tons, and jewelry.
Stationary versus non-stationary processes

Random processes may be characterized in terms of their temporal/spatial

And/or ensemble statistics

A random process is said to be stationary in the strict sense or strongly

stationary if its statistics are not affected by a shift in the origin of time or
space

In most practical applications, only the first-order and second-order

averages are used

A random process is said to be weakly stationary or stationary in the wide

sense if its mean is a constant and its ACF depends only upon the
difference (or shift) in time or space
Then we have

The ACF is now a function of the shift parameters α and β only; the PSD of the
process does not vary with space

A stationary process is said to be ergodic if the temporal statistics computed are

independent of the sample observed; that is, the same results are obtained with
any sample observation fk(t)

The time averages of the process are then independent of k:

All ensemble statistics may be replaced by temporal statistics when
analyzing ergodic processes

Ergodic processes are an important type of stationary random processes

because their statistics may be computed from a single observation as a
function of time

Signals or processes that do not meet the conditions described above

may, in general be called non-stationary processes

A non-stationary process possesses statistics that vary with time or space

The statistics of most images vary over space, indeed, such variations are
the source of pictorial information
Most biomedical systems are dynamic systems and produce non-stationary signals
and images

However, a physical or physiological system has limitations in the rate at which it

can change its characteristics

This limitation facilitates the breaking of a signal into segments of short duration
(typically a few tens of milliseconds) over which the statistics of interest may be
assumed to remain constant

The signal is then referred to as a quasi-stationary process

Techniques designed for stationary signals may then be extended and applied to
non-stationary signals

Analysis of signals by this approach is known as short time analysis

The characteristics of the features in an image vary over relatively large scales of
space, statistical parameters within small regions of space, within an object, or
within an organ of a given type may be assumed to remain constant

The image may then be assumed to be block-wise stationary, which permits

sectioned or block-by-block processing or moving-window processing using
techniques designed for stationary processes

~ - - ~ -

Ii
~

FIGURE 3.11
Block-by-block processing of an image. Statistics computed by using the pbcels
within the window shown with solid lines (3 x 3 pi.xels) are applicable to the
pi.xel marked with the @ symbol. Statistics for use when processing the pixel
marked with the# symbol (5 x 5 pixels) are computed by using the pbcels
within the window shown with dashed lines.
Certain systems, such as the cardiac system, normally perform rhythmic
operations

Considering the dynamics of the cardiac system, it is obvious that the

system is non-stationary

However, various phases of the cardiac cycle as well as the related

components of the associated electrocardiogram (ECG)
phonocardiogram (PCG) and carotid pulse signals - repeat over time in
an almost-periodic manner

A given phase of the process or signal possesses statistics that vary from
those of the other phases, however, the statistics of a specific phase
repeat cyclically
For example, the statistics of the PCG signal vary within the duration of a
cardiac cycle, especially when murmurs are present, but repeat
themselves at regular intervals over successive cardiac cycles

Such signals are referred to as cyclo-stationary signals

The cyclical repetition of the process facilitates synchronized ensemble

averaging using epochs or events extracted from an observation of the
signal over many cycles

The cyclical nature of cardiac activity may be exploited for synchronized

averaging to reduce noise and improve the SNR of the ECG and PCG

same technique may also be extended to imaging the heart: In gated blood-
pool imaging, nuclear medicine images of the heart are acquired in several
parts over short intervals of time. Images acquired at the same phases of the
cardiac cycle determined by using the ECG signal as a reference, trigger, or
"gating" signal are accumulated over several cardiac cycles. A sequence of
such gated and averaged frames over a full cardiac cycle may then be played
as a video or a movie to visualize the time-varying size and contents of the
left ventricle. (See Section 3.10 for illustration of gated blood-pool imaging.)
 Covariance and cross-correlation

vVhen two randorn processes f and g need to be corupared, we could cornpute

the covariance between thern as

where p119 (f, g) is the joint PDF of the two processes, and the irnage coor-
dinates have been ornitted for the sake of corupact notation. The covariance
pararneter rnay be norrnalized to get the correlation coefficient, defined as

\Vith -1 < PJg < +I.

A high covariance indicates that the two processes have similar statistical
variability or behavior

The processes f and g are said to be uncorrelated if ρfg = 0

Two processes that are statistically independent are also uncorrelated, the
converse of this property is, in general, not true

vVhen dealing with randoru processes f and g that are functions of space,
the cross-correlation function (CCF) between thern is defined as

~ f 9 (a, [3) - Elf (x, y) g( x + a, y +{3)].

Signal-dependent noise

Noise may be categorized as being independent of the signal of interest if

no statistical parameter of any order of the noise process is a function of
the signal

Although it is common to assume that the noise present in a signal or

image is statistically independent of the true signal (or image) of interest,
several cases exist in biomedical imaging where this assumption is not
valid and the noise is functionally related to or dependent upon the signal

Poisson noise

Imaging systems that operate in low-light conditions or in low-dose

radiation conditions such as nuclear medicine imaging, are often affected
by photon noise that can be modeled as a Poisson process
The probabilistic description of an observed image (pixel value) g0(m,n) under
the conditions of a Poisson process is given by

l;\ f(m,n)] 90
expl AJ(rri, n)]
(m,n)
P(go(m,n)lf(m,n), A)- ( )' , (124)
90 rn, n .

where f(m, n) is the undegraded pixel value (the observation in the absence
of any noise), and ;\ is a proportionality factor. Because the rnean of the
degraded image 90 is given by

Elgo(m, n)] - AElf (m, n)], (3.25)

irnages corrupted with Poisson noise are usually norrnalized as

( ) 9o(m,n)
g rri, n - ;\ . (3 .26)
Film-grain noise
The granular structure of film due to the silver-halide grains used
contributes noise to the recorded image, which is known as film-grain noise

When images recorded on photographic film are digitized in order to be

processed by a digital computer, film-grain noise is a significant source of
degradation of the information

The model for an image corrupted by film-grain noise is given by

g(m,n)- f(m,n) +l\.rlf(rn,n)]'ry1(rn,n) +ry2(rri,n), (128)

where K. is a proportionality factor, Fl ] is a rnathematical function, and

''lt (m, n) and T/2 (1n, n) are sarnples from two randorn processes independent of
the signal. This rnodel may be taken to represent a general in1aging situation
that includes signal-independent noise as well as signal-dependent noise, and
the noise could be additive or multiplicative. Observe that the rnodel reduces
to the sirnple signal-independent additive noise rnodel in Equation J.4 if K - 0.
 Froehlich et a1. ll81] ruodeledfilru-grain noise with Flf(rn, n)] - lf(m,n)JP,
using p - 0.5. The two noise processes T/t and T/2 were assurued to be Gaussian-
distributed, uncorrelated, zero-ruean randon1 processes. According to this
n1odel, the noise that corrupts the iruage has two coruponents: one that is
signal-dependent through the factor "' ✓ f (m, n) ·ry 1 ('m, n), and another that
is signal-independent given by ·ry2 (rn, n). Filn1-grain noise rnay be n1odeled
as additive noise as in Equation 14, \vith 'ry(rn,n)-"' JJ(ni,n) TJi(m,n) +
T/2 (m, n). It can be shown that TJ(ni, n) as above is stationary, has zero ruean,
and has its variance given by l182]

(:t29)
The fact that the rnean of the corrupted in1age equals the rnean of the noise-
free image has been used in an·iving at the relationship above.
Speckle noise

Speckle noise corrupts images that are obtained by coherent radiation, such as
synthetic-aperture radar (SAR), ultrasound, laser and sonar

When an object being scanned by a laser beam has random surface roughness
with details of the order of the wavelength of the laser the imaging system will be
unable to resolve the details of the object’s roughness – this results in speckle
noise

The most widely used model for speckle noise is a multiplicative model, given as

g (rn, n ) - f ('m, n ) 771 (m, n ) , (3 .30)

where 171 (rn, n) is a stationary noise process that is assurned to be uncorrelated

with the irnage. If the rnean of the noise process µ111 is not equal to one, the
noisy irnage may be norrnalized by dividing by µ111 such that, in the norn1alized
image, the multiplicative noise has its mean equal to one.
Depending upon the specific application, the distribution of the noise
may be assumed to be exponential, Gaussian, or Rayleigh

The multiplicative model in Equation 3.30 may be converted to the

additive model as in Equation 3.4 with η(m.n) being zero-mean additive
noise having a space-variant, signal-dependent variance given by

(]'2
l'T2(m
v n , ·n) _ t11 lag2( rn, n ) + µg2( m, n )] .
·, 1 + a t]1
2

In the expression above, a~ ( rn, n) and µ 9 ( rn, n) are the variance and the rnean
of the noisy irnage at the point (rn, n ), respectively.
 Transfurrnatiun of signal-dependent noise tu signal-independent
noise: In the rnodel used by Naderi and Sawchuk ll82J and Arsenault et al.
l189, 190], the signal-independent coruponent of the noise as in Equation 128
is assurned to be zero. In this case, it has been shown ll89, 190, 191] that by
applying an appropriate transforrnation to the whole irnage, the noise can be
ruade signal-independent. One of the transforrnations proposed is ll89, 190]

T lg (m, n) J- a / g(m, n),

where a is an appropriate norn1alizing constant. It has been shown that the

noise in the transforrned iruage is additive, has a Gaussian distribution, is
unbiased, and has a standcu-d deviation that no longer depends on the signal
but is given by 0/ l .
Synchronized or Multiframe Averaging
In certain applications of imaging, if the object being imaged can remain
free from motion or change of any kind (internal or external) over a long
period of time compared to the time required to record an image, it
becomes possible to acquire several frames of images of the object in
precisely the same state or condition

Then, the frames may be averaged to reduce noise, this is known as

multiframe averaging

The method may be extended to the imaging of dynamic systems whose

movements follow a rhythm or cycle with phases that can be determined by
another signal such as the cardiac system whose phases of contraction are
indicated by the ECG signal

Then, several image frames may be acquired at the same phase of the
rhythmic movement over successive cycles, and averaged to reduce noise
Such a process is known as synchronized averaging
The process may be repeated or triggered at every phase of interest
Synchronized averaging is a useful technique in the acquisition of several
biomedical signals

Let us represent a single image frame in a situation as above as

9i ( X, Y) - f ( X, Y) + T]i ( X, y), (3 .33)

where 9i(x, y) is the i th observed frame of the image f(x, y), and ·q1 (x, y) is the
noise in the same frame.

Let us assume that the noise process is independent of the signal source

Observe that the desired (original) image f(x, y) is invariant from one
frame to another
2
It_follo\vs_ that
_
0"
~i(x,y)
That is, the variance at every pixel in the observed noisy image is equal to
the corresponding variance of the noise process
 If lv[ frames of the irnage are acquired and averaged, the averaged image is
given by
1 M
g(x, y) - M L 9i(x, y).
i-1

If the mean of the noise process is zero, we have "2:t1

1 T/i ( x, y) ➔ 0 as M ➔ oo
(in practice, as the nurnber of frames averaged increases to a large nurnber).
Then, it follows that l8]
Elg(x,y)] - f(x,y), (3.35)

and
2 1 2
a §(x,y) - M a 17(x,y). (:1.36)

Thus, the variance at every pi.xel in the averaged in1age is reduced by a factor
of 1~ fron1 that in a single frarne; the SNR, is in1proved by the factor /M.
The most important requirement in this procedure is that the frames being averaged
be mutually synchronized, aligned, or registered
Any motion change, or displacement between the frames will lead to smearing and
distortion
 l'I t;UJU~ 3.12
(,) 11 ~1L11>:·1 : 1 12~, I!~ lt1I im,;r with v1ri m1 0 -auwtri:-·11 ubjtd1 pl 1:-t-:I
·,I rwbm. (bl lr111;ri11(·1J v.itltli nHmn 1wi-.• Tlcl.-:1, witlq1 11, J! l.'11
(11 :irunliu-:l), IU1b rrrn IlH. (r) "<-:1(11d vmi:m ~I 1.:1i1y im,;r, IU1t~
mm IY.it. llt11tll al 1rnutiln111r wtr 1;ilt; 1ui11; (1ll tkr lw.i lrimt1 in (bl
111d(<'), IUl~rrrur IUI; (r)lmr lr11tw1,llM~,'fr'rr l!.51; (l)c-i;lit
lrnm•1, ll.\1 ~ c-rr 1r I'.I.Yt

(n) (b j
Space-domain Local-statistics-based Filters

Practical situation - when we are given a single, noisy observation of an

image of finite size  Multiframe averaging is not possible

In such situations, moving-window filtering using windows of small size

such as 3X3, 5X5 or 7x7 are valuable options

Rectangular windows as well as windows of other shapes may also be

considered where appropriate

Various statistical parameters of the pixels within such a moving window

may be computed, with the result being applied to the pixel in the output
image at the same location where the window is placed (centered on the
input image
 input image output image

- - --
-- - ---"' - : :•• ii
u
'.al
.. ---
~

- -
~ - - -
~- '
i'@

,,

FIGURE 3.13
l\lloving-window filtering of an image. The size of the rnoving windo,v in the
illustration is 5 x 5 pixels. Statistics cornputed by using the pixels within the
window are applied to the pixel at the same location in the output image.
The moving window is shown for two pixel locations rnarked # and @.
 (a) 3X3 square (b) 4-conn.ect:ed tel 3x:. bar (d) :.x3 ba::-
( B-connecced) or integer
distance l.

(e) sx.s square (f) cross {g} sx:. bar ( t:) :.xs ba::-

(i) circl.e (j) incege::- (k) x-:. ( 1) X-2

distance 2

FIGURE :.S.14
A few con.unonly used [ll()~~in~-..vindow nei~hborhood shapes for irna~ filtcr-
in~. rhe result computed b_y usin~ the pL~cls within a window is applied to
the pL~cl at the location of its cent.er. sho._,,n shaded. in the output i.rua~c.
 The mean filter
If we were to select the pixels in a small neighborhood around the pixel to be
processed, the following assumptions may be rnade:

• the image component is relatively constant; that is, the irnage is quasis-
ta tionary; and

• the only variations in the neighborhood are due to noise.

Further assumptions regarding the noise process that are typically made are
that it is additive, is independent of the image, and has zero mean

Then, if we were to take the mean of the pixels in the neighborhood, the
result will tend toward the true pixel value in the original, uncorrupted
image

In essence, a spatial collection of pixels around the pixel being processed is

substituted for an ensemble of pixels at the same location from multiple
frames in the averaging process, that is, the image-generating process is
assumed to be ergodic
 -
It is common to use a 3 x 3 or 8-connected neighborhood as in Figure 3.14
( a) for mean filtering. Then, the output of the filter g( m, n) is given by

1 1 1
g(m,n)-
9
L L f(1n+a,n+f3),
a:- 1 {3- 1

where f(m,n) is the input image. The summation above may be expanded
as
1
g(rn, n) - x
9
l f (rn · 1, n 1) +f (m 1, n) + f (rn • 1, n + 1)
+f (rn, n - 1) +f (m, n) + f (rn, n + 1) (3.38)
+f(rn+ l,n-1) +f(m+ l,n) +f(rn+ l,n+ 1) ] .
The same result is also achieved via convolution of the irnage f (m, n) with
the 3 x 3 array or rnask
1 1 1
1
1 1 1 (3.39)
g
1 1 1
Note that the operation above cannot be directly applied at the edges of
the input image array

It is common to extend the input array with a border of zero-valued pixels

to permit filtering of the pixels at the edges

One may also elect not to process the pixels at the edges, or to replace
them with the average of the available neighbors

The mean filter can suppress Gaussian and uniformly distributed noise
effectively in relatively homogeneous areas of an image

However, the operation leads to blurring at the edges of the objects in the
image, and also to the loss of fine details and texture
 The median filter
The median of a collection of samples is the value that splits the population
in half

Half the number of pixels in the collection will have values less than the
median and half will have values greater than the median

In small populations of pixels under the constraint that the result be an

integer approximations will have to be made

The most common procedure rank-orders the pixels in a neighborhood

containing an odd number of pixels, and the pixel value at the middle of the
list is selected as the median

The procedure also permits the application of order-statistic filters

The median filter is an order-statistic filter of order N/2, where N is the size
of the filter
The median filter is a nonlinear filter

Its success in filtering depends upon the number of the samples used to
derive the output, as well as the spatial configuration of the
neighborhood used to select the samples

The median filter provides better noise removal than the mean filter
without blurring, especially when the noise has a long-tailed PDF
(resulting in outliers) and in the case of salt-and-pepper noise

However, the median filter could result in the clipping of corners and
distortion of the shape of sharp-edged objects

Median filtering with large neighborhoods could also result in the

complete elimination of small objects

Neighborhoods that are not square in shape are often used for median
filtering in order to limit the clipping of corners and other types of
distortion of shape
 Examples: Figure 3.15 (a) shows a 1D test signal with a rectangular pulse;
part (b) of the same figure sho\.vs the test signal degraded \.vith impulse (shot)
noise. The results of filtering the noisy signal using the rnean and median with
filter length N - 3 are sho\.vn in plots (c) and (d), respectively, of Figure 3.15.
The mean filter has blurred the edges of the pulse; it has also created artifacts
in the form of small hills and valleys. The median filter has removed the noise
without distorting the signal.

•wr
!'~~l~Ll~~~-~-~-~·-~~~u~,-~·
10 20 30 40 SO 60 70 80

l:~~t-~·--~l--~·--~·--~·--~~~--·~-~·
10 20 30 40
index
SO 60 70 80

FIGURE 3.15
( a) A lD test signal with a rectangular pulse. (b) Degraded signal with
impulse or shot noise. Result of filtering the degraded signal using (c) the
mean and ( cl) the median operation with a sliding window of N - 3 samples.
 Figure 3.16 (a) shows a ID test signal with two rectangular pulses; part (b)
of the same figure shows the test signal degraded with uniformly distributed
noise. The results of filtering the noisy signal using the mean and median with
filter length N - 5 are shown in plots (c) and (d), respectively, of Figure 3.16.
The rnean filter has reduced the noise level, but has also blurred the edges of
the pulses; in addition, the strength of the first, short pulse has been reduced.
The median filter has removed the noise to some e..-x:tent without distorting
the edges of the long pulse; however, the short pulse has been obliterated.

150~

t·:t~~D~.-~l-~.-~·-~\-~·~·
10

150~ {\
20 30 40 50 60 70 80

r:~ 10 20 30 40 50 60

10 20 30 40 50 60 70 80

l:~LL,
10 20 30 40 ~
50 60 -
70 80
index

FIGURE 3.16
( a) A lD test signal with two rectangular pulses. (b) Degraded signal with uni-
formly distributed noise. Result of filtering the degraded signal using ( c) the
mean, and (d) the median operation with a sliding window of N - 5 samples.
 Figure 3.17 sho,vs the original test image "Shapes", the test image degraded
by the addition of Gaussian-distributed random noise ,vith µ - 0 and a- 2 -
0.01 (normalized), and the results of filtering the noisy image ,vith the 3 x 3
and 5 x 5 mean and median filters. The RMS errors of the noisy and filtered
images with respect to the test image are given in the figure caption. All of
the filters except the 3 x 3 median have led to an increase in the RlVIS error.

The blurring effect of the n1ean filter is readily seen in the results. Close
observation of the result of 3 x :1 rnedian filtering lFigure :1.17 (d)] sho\vs that
the filter has resulted in distortion of the shapes, in particular, clipping of
the corners of the objects. The 5 x 5 rnedian filter has led to the con1plete
rernoval of srnall objects; see Figure 3.17 (f). Observe that the results of the
J x 3 rnean and 5 x 5 rnedian filters have sirnilar Rl\lIS error values; however,
the blurTing effect in the forrner case, and the distortion of shape information
as well as the loss of srnall objects in the latter case need to be considered
carefully.
fl GUttK 3.1,
j·,1 ""1-,pn lt-11 tun;:.!. [b) l.J:1:1:::c, Ht 1,:1 with G;m1t11nn 1wi-.t• .-m,.lNi. wllb
,, .I. tr - J .,11 'n Ir IU ,Ii z,-,l .1. tu,1 .. MT"-'J" rn. '>li. ltt'1 nll <.I I fi It t'I" D>;; I h-
11><1 uy 1m:,~ in (1,) 1uin:;: (c) .1 { Wf",:m, tL\t~t-rror .U.{L!; (.l I ,;l, :{ mt"diro:.
IU,l, •rr,1r l;",.J.I: ,,., ; • ; m<c".:Ln. ll.'•L' <.'nor .:!~/I"; ( f l ; ; m~li:m.
H ..{5.
 k) fd)

It"} (I )

H(;Ultl!: 3.18
111 'l-h:ip,•1 lt":Sl 1m-''>"· fb,1 lm-,;.,. m (n.l v,;ll, 1'<111-...1111.uix-. ll.\i"'('D"Ur 'i. .lU.
ltf:fnll JI Wl<-J"ir:.3 Uit- n<ti.y iru.:\!,e iu I bJ ,~m:;: jr) 3-<. j mi-.:>1t. ll.\1!) e,rrnr
rn . .;.I; fd} "iv 1 med.L,n. lt:,t, errnr l~.19; fe) 5 • 5 me-..:ut, lt..\t.'- <•rr·:u
.fi.'!5; (I) fl • ; w:-ctnu, lt:~l'- crrur .LLfi.
 in) (b)

ki (cl)

(<"I if)

t'IGUlt~ !U9
h.l ,t..,pc, lt<1I im:i,.c-. (b} Lm:,;:t< m h) with ·IJ>C't:kJ!' faJNt'. with I' I, 'T
O.'.lt ln.rrm'.lb~). IL\t'!-c-rrm I.L~'I. ltt"11i.ll ,I lillt-I11-R tllll"nw,ymn;<'in
lb) ,um;;: l<-.1 j ,. j m,,n. tt:,t'- •nm !J.'.i I; (<l} :i, :i m~.:m. IL\{') c-.nor
1;.Ci1i: (t!) 5 • 5 mc;iu. ltM'- err ir !ti.:U; 1n 5 ·, 5 mt'W,n. ll.\t."> ~rnr
!1.:lli.
 .
..... -
.~-- : •.
...
.:.. , .. :
r.-.
L • • .• . . U

(:-i) (b)

( t') (1)

Jll(;UH.l!. 3.20
(:-i) 'i,h.."J~"" lt-11 irw,,;t'. (b) lm:i;,'t' m ('t) with -.:ill ·md JM"J>J>t'r uui1t' .,d.<Lt-:l.
wllh dc-1u1l:,- ;1.:15. IL\t') t"rror t 1.:.1:1. ltc-'full ol lillt'.rm;; tlllt' llW'i)' 1m't;;c-
111 (b) 1um;;: (<') :! • i mc-.'tn. H.~•t'l- t'rrur .H.'!5: (d) :! "·:1 mt'd1'\11. te,t.'l- c-rror
I t.5J: (c-) 5 , 5 mt':tn. ll.\t'l- c-rrur .:!'! . .:!~; (I) 5 , 5 mt-:Ls,n. 1-L\l.'l, c-rnr
.:!Ul.
 •
• ~

•
131
~
•• . a
b. II •
-- ~ " ~ -
1"1GUH.~ 3.21
h) '-lt.:>pe, lte,I uu,;;c•. (h) Im·•~ it. f,) will. 'l.'111 ~m! J>:'J>J,-..r 11<.ci~~ '1<klt<CI.
v.1lh cl<-inity 11.1. IL\t, .-rrc,r 5li.'.i!. ltc-:mll JI Wknn:; the u,ony iru,;;c-1u
lb) llc'IH~: [c·.1 .i • :.i mt-:,n, llM'- .-rnr !!UH: Jcl\ i-. :l m~,1t, ltM~ rrru1
15 .H; It-) :. ~ 'i 1t1¢<~11t. IC.t" c-rrc.cr :ll ..fi; I I) 'i ~ :i lllf'(U ·m, ltMS crr-ir
1~.:-11.
 I I
(a) ( b)

(c) (d)

FIGGRE 3.22
:Xormalized histo~ram'i of (a) the ShaJ)l's test ima~l' and of lhl' ima~e wilh
(b) Catis'iia11 noise. (c) Poisson noise. and (d} spl'ckll' noisr. fhe first his-
to~ram has brrn scaled t.o display thr ran~ of probability (tJ. IJ.05) only: lhl'
remainin~ hist.o~ams ha,·l' bern scalrd t.o display lhl' ran~l' (0. IJ.(Jl5) only
in ordrr t.o show lhc important details. The probability \'Blur'i of ~ray k\'-
els tJ and 255 ha\'C brrn clipped in some of lhc hist.o~rams. Each hi'ito~ram
reprl'srnts lhe ~ay-k\'rl ran~c of 0. 255 .
Order-statistic filters
The class of order-statistic filters is large and includes several nonlinear
filters that are useful in filtering different types of noise in images

The first step in order-statistic filtering is to rank-order from the minimum

to the maximum, the pixel values in an appropriate neighborhood of the
pixel being processed

The ith entry in the list is the output of the ith order-statistic filter

A few order-statistic filters of particular interest are the following:

• _Min fill er: the first entry in the rank-ordered list, usc!ul in reruoving
high-valued impulse noise (isolated b1·ight spots or lLsalt=' noise).

• _Ma2 fi1le1·: the last entry in the 1·ank-ordered list, useful in reruoving
low-valued irupulse noise (isolated dark spots or upcppei·=' noise).

• _n!Jin/-~11'1.2 filte?': sequential avplication of the wtio and Nta..x. filtei·s, useful
in ren1oviog salt-and-pevpe1· noise.
Median filter: The entry in the middle of the list.

The median filter is the most popular and commonly used filter
among the order-statistic filters

α-trimmed mean filter: The mean of a reduced list where the first α
and the last α of the list is rejected, with 0 < α < 0.5

Outliers, that is pixels with values very different from the rest of the
pixels in the list are rejected by the trimming process

A value close to 0.5 for α rejects the entire list except the median or a
few values close to it and the output is close to or equal to that of the
median filter

The mean of the trimmed list provides a compromise between the

generic mean and median filters
L-Filters: A weighted combination of all of the elements in the rank-
ordered list

The use of appropriate weights can provide outputs equal to those of all of
the filters listed above, and facilitate the design of several order-statistic
based nonlinear filters

Order-statistic filters represent a family of nonlinear filters that have

gained popularity in image processing due to their characteristics of
removing several types of noise without blurring edges, and due to their
simple implementation
Frequency-domain Filters
Transforming an image from the space domain to the frequency domain
using a transform such as the Fourier transform provides advantages in
filtering and noise removal

Most images of natural beings, entities, and scenes vary slowly and
smoothly across space, and are usually devoid of step-like changes

As a consequence, such images have most of their energy concentrated

in small regions around (u, v) = (0, 0) in their spectra

On the other hand, uncorrelated random noise fields have a uniform,

flat, or “white” spectrum, with an almost-constant energy level across
the entire frequency space

This leads to the common observation that the SNR of a noisy, natural
image is higher in low-frequency regions than in high-frequency regions
 -
The procedure for Fourier-domain filtering of an irnage f (rri, n) involves the
following steps:

1. Cornpute the 2D Fourier transforrn F(k, l) of the irnage. This rnay

require padding the irnage with zeros to increase its size to an N x N
array, with N being an integral power of 2, if an FFT algorithm is to
be used.

2. Design or select an appropriate 2D filter transfer function H( k, l).

J. Obtain the filtered image (in the Fourier dornain) as

G(k, l) - H(k,l) F(k,l). (3.40)

It is common to define H(k, l) as a real function, thereby affecting only
the magnitude of the input image spectrum; the phase remains
unchanged

4. Corupute the inverse Fourier transforrn of G(k, l). If F(k, l) was folded
prior to filtering, it rnust be unfolded prior to the inverse transforrnation.

5. If the input in1age was zero-padded, trin1 the resulting in1age g(rn, n).
Removal of high-frequency noise

Low-pass filters are useful in removing high-frequency noise, under the

assumption that the noise is additive and that the Fourier components of
the original image past a certain frequency cutoff are negligible
The so called ideal low-pass filter is defined in the 2D Fourier space as

H {lifD(u,v)<Do
( u, v) - 0 other\vise.

where D( u, v) - Ju 2 + v2 is the distance of the frequency cornponent at

(u,v) frorn the DC point (tt,v) - (0,0), with the spectrurn being positioned
such that the DC is at its center
Note: The coordinates (u, v) and (k, l) are used interchangeably to
represent the 2D frequency space
D0 is the cutoff frequency, beyond which all components of the Fourier transform
of the given image are set to zero

Figure 3.28 a shows the ideal low-pass filter function

(a) (b)

FIGURE 3.28
( a) The magnitude transfer function of an ideal lowpass filter. The cutoff
frequency Do is 0.4 times the maximum frequency (that is, 0.2 times the
sampling frequency). (b) The magnitude transfer function of a Butterworth
lowpass filter, with normalized cutoff Do - 0.4 and order n = 2. The (u,v) =
(0, 0) point is at the center. The gain is proportional to the brightness ( white
represents 1.0 and black represents 0.0.)
Figure 3.29 shows profiles of the ideal and Butterworth low-pass filters

1 ........................ . ·········-~-·-~~~
,i.... .... ~

/
/ :
:
:
:
' '\
: '
........ +·. .....',. ·········~······
I
0.8 ......... ·I·······~···· ....
.. ... .. .
.... ....
C:
.; . . I . . .: I .: .
Cl
;;; 0.6 .. . . . . . ..: ......... .: ...... / ............~ ..........:...........:...........\ ...... ~.......... ;........ .
= .. .. .. .. .. .. ..
u: : : I : : : : :
:
: :
:
'' :
:
:
: :
: ' :: :
:
........ :.........:. /. ...............~ .........:..........:............... ~.. ;......... :....... .
0.4 : : 1 : : : 1: :
: j, : • • ,: :
: ,: : :., :
:: / / :: :: :
:: ' ' ::
02 ········:···~····:·········, ·········~········ .......• .:••••••••• ~•••• ~ ... ; ....... .
:o/1 / : o
,: :I o
' :
,•

., , "':.' :.. :.' :'

.. .. ..
~.5 -0.4 -0.3 -02 -0.1 0 0.1 0.2 0.3 0.4 O.S
Nonnalized frequency

FIGURE 3.29
t'rofiles of the magnitude transfer functjons of an ideal lowpass filter (solid
line) and a But.terworth lowpass filter (dashed line), with norruali;-;ed cutoff
Do - 0.4 and order n - 2.
 (a) (b)

(cl (d)

FIGGRE :i.:m
(a) fhc Shaprs lc:st. image. (b) The t.esl inuu;c \7,;u, Gaus..ian noise: ha,;ni;
a nonnaliurl varianct.> of 0.fJl arldcrl. (c) Thl' result of idral lowpass filtrrin~
the uoisy ima~e. wiLh normalucd cutoff Du UA: scr Fi~rr :l.28. (d) fhe
result of filtering with a Ouuerworth lowpass filler ha,·in i; D. 1 U. I anti order
n 2. ~re also Fii;urc 3.:!l.
 (b I

FIGURE a.al
l'he r~nterrd ~ fold(l"d) Fourier lo~-nu~~itude spcct.ruw of I. al I he ~hapes 1ru-
a~cs in Fi~ure :t:~u (P1J aud (bl Lhe noifi~• ~lu~ptis ima~e in Fi~urr :i.:w (b).
Multiplication of the Fourier transform of the image with the circle
function is equivalent to convolution of the image in the space
domain with the corresponding Bessel function

The ripples or lobes of the Bessel function lead to echoes of strong

edges, an artifact known as the ringing artifact

The example illustrates that the ”ideal” filter’s abrupt transition

from the passband to the stopband is, after all, not a desirable
characteristic
 The Butterworth lowpass filter

Prevention of the ringing artifacts encountered with the ideal lowpass filter
requires that the transition from the passband to the stopband (and
viceversa in the case of highpass filters) be smooth

The Butterworth filter is a commonly used frequency domain filter due to its
simplicity of design and the property of a maximally flat magnitude
response in the passband

For a 2D Butterworth lowpass filter of order n, the first (2n-1) derivatives of

the squared magnitude response are zero at ω = 0, where ω represents the
radian frequency

The Butterworth filter response is monotonic in the passband as well as in

the stopband
 In 2D, the Butterworth lowpass filter is defined as l8]

(142)

where n is the order of the filter, D(u, v) - \/u2 +v2, and Do is the half-
power 2D radial cutoff frequency lthe scale factor in the denorninator leads to
the gain of the filter being ~ at D( u, v) - Do]. The filter's transition from
the passband to the stopband becornes steeper as the order n is increased.
Figures 128 (b) and 129 illustrate the r11agnitude (gain) of the Butterworth
lowpass filter with the norrnalized cutoff D0 - 0.4 and order n - 2.
 •I
'
t

- '
.----- ......... -- ---- •

ft
'
I
•'
I

(aJ (bl

(c) ( d)

FIG GUE :i. J2

(a) Clock test irua~e (Hll A 101 pixels). (b) Lo~-ma~uitudc spccl.rum of I.he
ima~c. (c) R=ulL of Lbr i<leal lowpass fihcr. DJ UA. (d) Resuh of U1c
D11U,crworib lowpass fill.er. with D 0 U.4 and orrlcr n 2.
Removal of periodic artifacts

Periodic components in images give rise to impulse-like and periodic

concentrations of energy in their Fourier spectra

This characteristic facilitates the removal of periodic artifacts through

selective band-reject, notch or comb filtering

Example: Figure 3.33 (a) shows a part of au iwage of a ruaruwographic

phautoru acquired with the grid (bucky) reruaiuiug fixed. The white objects
siwulate calcifications fouud in waU1Wog1·aws. The projections of the g1·id
strips have suppressed the details in the phautow. Figtuc 3.33 (b) shows the
log-ruaguitude Fourier spectruw of the iruage, where the periodic couceutra-
tious of eue1·gy aloug the v ax.is as well as along the correspoudiug horiliontal
strips are related to the grid lines. l·•'igure 3. 33 (d) shows the spectruru with
selected regions corrcspouding to the artifactual coUtpoucnts set to liero. Fig-
ure 3.33 (c) sho"vs the corresponding filtered iruage, where the g1·id lines have
been alUtost corupletely reruoved.
 - - - ---
- --- ------ --- - -
- --- - -
-- - ---
-- - - - --
- -- -- - -
-
---
- - - - - - --- - -
- - -- -

-t

k) fd)

1-'J t:;U H. It 3. 33
(.11 l'.:i.rl oJ .:,u i11J'l:;c- <1.I ,. 111;m1ma;rapld:- ph.:lnl<IW witL gdri ,rUl.,,;·l; :s~·
.i.l:s,J l· j;;1tre .!L:~.2. ib.l L<1g mri:;.r..iu~<r..t" l-<11:.ri~.r 'fJ>N"l-rmu c,J tlw- 11:.1.:a~t: in (:ij.
k) J:- illt:ft~d jma;~. id) J:-1Jl<•rec.l ,rt-.rsic111 a.I lb:- :SJHx·lrum iu. fl>). i'h.:iulum
iluiJg<' cmcrl~• -if L..J. H:thn.. l·<mUdll\ l1C1,1pit;.il. C.:i_J-;:uy.
 Figur" :1.:~4 la) . . h ,•, a oorr "'prnuling ituage of the phant,0111 acC)nirn l •.1rith
1: : i 1

the bu k,· 1u0Yin~ 1u the reco1111nendc-<l 111anner; the itua~e i:-. free of th· grid-
line artifact. Fi~ure :L:~4 (b) "'h'J',,,r-; the correspon lin~ log-1ua~nit ucle l•bnrier
1

:'11,oct nun. •.vhich L, aL"'o free oft h:, art ifRct ual con!p nient ~ that are :-.ee u in the
:'lpect nun in Fi~ure :L:i:~ (bl. It. :'lhoul l b noted t h;.t. re1110Ying the artifact ual
co1np Jnenl-;; a:; ind1ust~d by th :-ipect nun in Figure :3. :~:i (d) lead;; tc, the lo5:i
oft h .. fr,eqtll?nc~·-do1uaiu con1pCJnent :-. ;_if th:, de:'lired ituage in the :-iRLU£ regitJLb.
which could lead to '501ue dhtiOrtion int he filtered ituage.

(a) (b I

FIGL'RE 3.54
( a) Part. of an imaN' of a mammo~raphic phautom \,.;th no ~rid artifact.; com-
pare wiU1 Lhc im~t! iu Fi~1.u:c :t:1:1 (a}. ( b) lo~-tn!lgnit.udc Fourier sprctnun
of I.hr iwa~e in (aJ. Phaut.om un!\,gc courtC"S}' of l..l. llarut. Foothills Dospital,
CaJ~at)'.
Applications

3.9 Multiframe averaging in confocal microscopy

3.10 Noise reduction in nuclear medicine imaging