2D Signals and Systems

Signals

• A signal can be either continuous f (x), f (x, y), f (x, y, z), f (x)
• or discrete f etc. where i,j,k index specific coordinates
i, j, k

• Digital images on computers are necessarily discrete

sets of data
• Each element, or bin, or voxel, represents some value,
either measured or calculated
Digital Images

• Real objects are continuous (at least above the quantum level), but we
represent them digitally as an approximation of the true continuous process
(pixels or voxels)
• For image representation this is usually fine (we can just use smaller voxels
as necessary)

• For data measurements the element size is critical (e.g. Shannon's sampling
theorem)
• For most of our work we will use continuous function theory for convenience,
but sometimes the discrete theory will be required
Important signals - rect() and sinc() functions

• 1D rect() and sinc() functions

– both have unit area

!1, for x < 1 / 2

(a) rect(x) = "
#0, for x > 1 / 2
sin($ x)
(b) sinc(x) =
$x
what is sinc(0)?
Important signals - 2D rect() and sinc() functions

• 2D rect() and sinc() functions are straightforward generalizations

!1, for x < 1 / 2 and y < 1 / 2

(a) rect(x, y) = "
#0, otherwise
sin($ x)sin($ y)
(b) sinc(x, y) =
$ 2 xy
• Try to sketch these
• 3D versions exist and are sometimes used
• Fundamental connection between rect() and sinc() functions and very
useful in signal and image processing
 Important signals - Impulse function
• 1D Impulse (delta) function
! (x) = 0, x " 0,
$
• A 'generalized function'
– operates through integration % ! (x) dx = 1
#$
– has zero width and unit area
– has important 'sifting' property $
– can be understood by considering:
%
#$
f (x)! (x) dx = f (0)
• Ways to approach the delta function $

%
#$
f (x)! (x # t) dx = f (t)

$ % a2t 2
! (t) = lim a rect(at) ! (t) = lim a sinc(at) ! (t) = lim ae
a"# a"# a"#
Exponential and sinusoidal signals

• Recall Euler's formula,which connects trigonometric and

complex exponential functions
e j! = cos(! ) + j sin(! ) (not i)

• The exponential signal is defined as:

e j 2 ! x = cos(2! x) + j sin(2! x), where j 2 = "1
• u0 and v0 are the fundamental frequencies in x- and y-
directions, with units of 1/distance e(x, y) = e j 2 ! (u0 x + v0 y)

j 2 ! (u0 x+v0 y)
• We can write e(x, y) = e
= cos "# 2! ( u0 x + v0 y ) $% + j sin "# 2! ( u0 x + v0 y ) $%

real and even imaginary and odd

Exponential and sinusoidal signals

• Recall that
sin(2! x) =
1 j 2! x
2j
e (
" e " j 2! x
)
1 j 2! x
cos(2! x) = e
2
(
+ e" j 2 ! x )

• so we have sin "# 2! ( u0 x + v0 y ) $% =

2j
e(
1 j 2 ! (u0 x+v0 y) & j 2 ! (u0 x+v0 y)
&e )
" $
2
(
1 j 2 ! (u0 x+v0 y) & j 2 ! (u0 x+v0 y)
cos # 2! ( u0 x + v0 y ) % = e +e )
• Fundamental frequencies u0, v0 affect the oscillations in x and y
directions, E.g. small values of u0 result in slow oscillations in the x-
direction
• These are complex-valued and directional plane waves
 Exponential and sinusoidal signals

• Intensity images for s(x, y) = sin "# 2! ( u0 x + v0 y ) $%

x
System models
• Systems analysis is a powerful tool to characterize and control the
behavior of biomedical imaging devices
• We will focus on the special class of continuous, linear, shift-
invariant (LSI) systems
• Many (all) biomedical imaging systems are not really any of the
three, but it can be useful tool, as long as we understand the errors
in our approximation
• "all models are wrong, but some are useful" -George E. P. Box
• Continuous systems convert a continuous input to a continuous
output

g(x) = S [ f (x)] ( g(t) = S [ f (t)])

f (x) S g(x)
Linear Systems

• A system S is a linear system if: we have S [ f (x)] = g(x)

then S [ a f (x) + a
1 1 f (x)] = a1g1 (x) + a2 g2 (x)
2 2

S "K % K
S [ f (x)] = ! w g (x)
K

$ ! wk fk (x) ' = ! wk
or in general
k k k
# k =1 & k =1 k =1

• Which are linear systems? g(x) = e! f (x)

g(x) = f (x) + 1
g(x) = x f (x)
g(x) = ( f (x))
2
2D Linear Systems

• Now use 2D notation

• Example: sharpening filter

f (x, y) g(x, y)

• In general

S "K % K
S [ f (x, y)] = ! w g (x, y)
K

$ ! wk fk (x, y) ' = ! wk k k k
# k=1 & k=1 k=1
Shift-Invariant Systems
• Start by shifting the input f x0 y0 (x, y) ! f (x ! x0 , y ! y0 )

then if
gx0 y0 (x, y) = S !" f x0 y0 (x, y) #$ = g(x % x0 , y % y0 )
the system is shift-invariant, i.e. response does not depend on
location

• Shift-invariance is separate from linearity, a system can be

– shift-invariant and linear
– shift-invariant and non-linear
– shift-variant and linear
– shift-variant and non-linear
– (what else have we forgotten?)
Shift invariant and shift-variant system response

scanner
object shift
f (x, y)
S
image

g(x, y)
FOV
unshifted response shift invariant shift variant
(shape, location)
Shift invariant and shift-variant system response

scanner
object shift
f (x, y)
S
image

g(x, y)
FOV
unshifted response shift invariant shift variant
(value)
 Impulse Response
• Linear, shift-invariant (LSI) systems are the most useful
• First we start by looking at the response of a system using a point
source at location (ξ,η) as an input

point
object
input f!" (x, y) ! # (x $ !, y $ " )
y
x
output g!" (x, y) ! h(x, y; !, " )

• The output h() depends on location of the point source (ξ,η) and location
in the image (x,y), so it is a 4-D function
• Since the input is an impulse, the output is called the impulse response
function, or the point spread function (PSF) - why?
 Impulse Response of Linear Shift Invariant Systems

• For LSI systems S [ f (x ! x , y ! y )] = g(x ! x , y ! y )

0 0 0 0

• So the PSF is
S [! (x " x , y " y )] = h(x " x , y " y )
0 0 0 0

• Through something called the superposition integral, we can show that

$ $
g(x, y) = % % f (!, " )h(x, y; !, " ) d! d"
#$ #$

• And for LSI systems, this simplifies to:

$ $
g(x, y) = % % f (!, " )h(! # x, " # y) d! d"
#$ #$

• The last integral is a convolution integral, and can be written as

g(x, y) = f (x, y)! h(x, y) (or f (x, y)! !h(x, y))

Review of convolution
#
• Illustration of h(x) = f (x) ! g(x) = $"#
f (u)g(x " u)du

original functions

g(x-u), reversed and shifted to x

curve = product of f(u)g(x-u)

x

area = integral of f(u)g(x-u)

= value of h() at x
x
Properties of LSI Systems

• The convolution integral has the basic properties of

1. Linearity (definition of a LSI system)
2. Shift invariance (ditto)

3. Associativity g(x, y) = h2 (x, y) ! [ h1 (x, y) ! f (x, y)]

= [ h2 (x, y) ! h1 (x, y)] ! f (x, y)
4. Commutativity h1 (x, y) ! h2 (x, y) = h2 (x, y) ! h1 (x, y)

Equivalent
arrangements
Combined LSI Systems

• Parallel systems have property of

5. Distributivity

g(x, y) = h1 (x, y) ! f (x, y) + h2 (x, y) ! f (x, y)

= [ h1 (x, y) + h2 (x, y)] ! f (x, y)
 Summary of advantages of Linear Shift Invariant Systems

• For LSI systems we have f (x, y) h(x,y) g(x, y)

object system image

$ $
g(x, y) = % % f (!, " )h(! # x, " # y) d! d"
#$ #$

= f (x, y)&&h(x, y)
• Treating imaging systems as LSI significantly simplifies analysis
• In many cases of practical value, non-LSI systems can be approximated
as LSI
• Allows use of Fourier transform methods that accelerate computation
2D Fourier Transforms
Fourier Transforms
• Recall from the sifting property (with a change of variables)
% %

f (x, y) = &&
$% $%
f (!, " )# (! $ x, " $ y) d! d"

• Expresses f(x,y) as a weighted combination of shifted basis

functions, δ(x,y), also called the superposition principle
• An alternative and convenient set of basis functions are sinusoids,
which bring in the concept of frequency
• Using the complex exponential function allows for compact notation,
with u and v as the frequency variables

e j 2 ! (ux+vy) = cos "# 2! ( ux + vy ) $% + j sin "# 2! ( ux + vy ) $%

 Exponential and sinusoidal signals as basis functions

• Intensity images for s(x, y) = sin "# 2! ( u0 x + v0 y ) $%

x
Fourier Transforms
• Using this approach we write
# #

f (x, y) = $$
"# "#
F(u, v)e j 2 ! (ux+vy) du dv
• F(u,v) are the weights for each frequency, exp{ j2π(ux+vy)} are the
basis functions
• It can be shown that using exp{ j2π(ux+vy)} we can readily calculate
the needed weights by
# #

F(u, v) = $$
!# !#
f (x, y)e! j 2 " (ux+vy) dx dy

• This is the 2D Fourier Transform of f(x,y), and the first equation is

the inverse 2D Fourier Transform
Fourier Transforms
• For even more compact notation we use

F(u, v) = F2D { f (x, y)} , and f (x, y) = F2D-1 {F(u, v)}

• Notes on the Fourier transform
– F(u,v) can be calculated if f(x,y) is continuous, or has a finite number of
discontinuities, and is absolutely integrable
– (u,v) are the spatial frequencies
– F(u,v) is in general complex-valued, and is called the spectrum of f(x,y)
• As we will see, the Fourier transform allows consideration of an LSI
system for each separate sinusoidal frequency
Fourier Transform Example
rect(x,y)
• What is the Fourier transform of
y
!1, for x < 1 / 2 and y < 1 / 2
rect(x, y) = " x
# 0, otherwise
• First note that it is separable rect(x, y) = rect(x)rect(y)
F {rect(x)} = $ rect(x)e
#
• So we compute ! j 2 " ux
1D dx
!#
1/2
1 ! j 2 " ux 1/2
$
! j 2 " ux
= e dx = e
!1/2
j2" u !1/2

1 e j" u ! e! j" u sin(" u)

= =
"u j2 "u
= sinc(u)
Thus F {rect(x, y)} = sinc(u, v)
2D
Fourier Transform Example

F { f (x, y)} ! F(u, v)

2D

rect(x, y) sinc(u, v)
Two Key Properties of the 2D Fourier Transform

• Linearity F {a f (x, y) + a g(x, y)} = a F(u, v) + a G(u, v)

2D 1 2 1 2

• Scaling F { f (ax,by)} = 1 ! u v$
F# , &
ab " a b %
2D
Signal localization in image versus frequency space

more less
localized localized
Higher spatial frequencies

less more
localized localized
 Fourier Transforms and Convolution
• Very useful! F { f (x, y) ! g(x, y)} = F(u, v)G(u, v)
2D
• Proof (1-D)

F { f (x) ! g(x)} = % ( f (x) ! g(x)) e

$
" j 2 # ux
dx
"$

' $ $
* " j 2 # ux $
' $
*
= %) % f (& )g(x " & )d& , e dx = % f (& ) ) % g(x " & ) e " j 2 # ux
dx , d&
( "$ "$ + "$ ( "$ +

%F
$
' $
* $

= % f (& ) )
(
{g(x " & )}, d& =
+
% (
f (& ) e" j 2 # u& G(u) d& )
"$ "$ "$
$

= G(u) % f (& )e" j 2 # u& d& = F(u)G(u)

"$
 Fourier transform pairs

• Note the reciprocal symmetry in Fourier transform pairs

– often 2-D versions can be calculated from 1-D versions by seperability
– In general: a broad extent in one domain corresponds to a narrow extent in the other
domain
Summary of key properties of the Fourier Transform
Transfer Functions
Transfer Function for an LSI System

• Recall that for an LSI system f (x, y) S g(x, y)

% %
g(x, y) = f (x, y) ! h(x, y) = & &
$% $%
f (", # )h(" $ x, # $ y) d" d#
• We can define the Transfer Function as the 2D Fourier transform of
the PSF

h(!, " )e j 2 # (u! + v" ) d! d" = F2D {h(x, y)}

% %
H (u, v) = & &
$% $%

• In this case the LSI imaging system can be simply described by:

g(x, y) = f (x, y) ! h(x, y) = F2 D"1 {F(u, v)H (u, v)}

• or G(u, v) = F(u, v)H (u, v)

• which provides a very powerful tool for understanding systems

Illustration of transfer function f (x, y) h(x,y) g(x, y)

f (x, y) g(x, y)

2-D FT Inverse 2-D FT

F(u, v) G(u, v)
a1 a2 > a1

! " a 2 (u 2 + v 2 )
H (u, v) = ae
X-ray Radiography
Definitions

• Ion: an atom or molecule in which the total number of

electrons is not equal to the total number of protons,
giving it a net positive or negative electrical charge

• Radiation: a process in which energetic particles or

energetic waves travel through a medium or space
 Ionizing Radiation
+
• Radiation (such as high energy
electromagnetic photons behaving like
particles) that is capable of ejecting
orbital elections from atoms
• Can also be particles (e.g. electrons)
• Ionizing energy required is the binding
energy for that electron's shell
• Energy units are electron volts (eV or
keV), the energy of an electron
accelerated by 1 volt Energies for Tungsten (W)
• For Hydrogen K orbital electrons, E=13.6
eV
• For Tungsten K orbital electrons, E=69.5
keV
• In medical imaging we need photons with
enough energy to transmit through tissue
so are in range of 25 keV to 511 keV and
is thus ionizing
 Electrons as Ionizing Radiation
• Electron kinetic energy E = (mv 2 ) / 2
• Three main modes of interaction in
the energy range we are
considering
a) Collision with other electrons and
possible creation of delta-rays
(high-energy electrons)
– This is the most common mode and
excited atoms loose energy by IR
radiation (heat)
b) Ejection of an inner orbital electron
– This orbit is filled by an outer
electron and the difference in
energy is released as a
'characteristic x-ray'
c) Bending of trajectory by nucleus
– Since acceleration of a charged
particle causes radiation, this
causes 'braking radiation' or
bremsstrahlung
X-ray Spectrum from Electron Bombardment
When high energy electrons hit tungsten (symbol W), three effects occur
1. Heat (> 99.9% of the energy)
2. Characteristic x-rays e-
3. Bremsstrahlung x-rays W

ΔV
accelerating voltage

Energies for Tungsten (W)

69.081 keV

59.321 keV