31-01-13871 2d sampling goal: represent a 2d function by a finite set of points. - particularly...
TRANSCRIPT
31-01-1387 1
2D Sampling
Goal: Represent a 2D function by a finite set of points.
- particularly useful to analysis w/ computer operations.
Points are sampled every X in x, every Y in y.
How will the sampled function appear in the spatial frequency domain?
31-01-1387 2
Two Dimensional Sampling: Sampled function in freq. domain
How will the sampled function appear in the spatial frequency domain?
)G(Y)III(X)III(XY
),(g),(G
u,vvu
yxvu
F
Since
n m Y,
XδY)comb(X)comb(XY
mv
nuvu
The result: Replicated G(u,v), or “islands” every 1/X in u, and 1/Y in v.
n m Y,
XG),(G
mv
nuvu
31-01-1387 3
Example
Let g(x,y) =(x/16y/16)be a continuous function
Here we show its continuous transform G(u,v)
-n -m
)g(mY)-nX,-δ(
),(gY
IIIX
III),(g
x,yyx
yxyx
yx
Now sampling the function gives the following in the space domain
31-01-1387 4
v
u
Fourier Representation of a Sampled Image
1/X
1/Y
Sampling the image in the space domain causes replication in the frequency domain
31-01-1387 5
Two Dimensional Sampling: Restoration of original function
will filter out unwanted islands.Y)(X)(),(H vuvu
Let’s consider this in the image domain.
),(h),(g yxyx
Ysinc
Xsinc
XY
1**
)mY,nX(δ)mY,nX(gXY
),(h),(gY
IIIX
III
n m
yx
yx
yxyxyx
31-01-1387 6
Two Dimensional Sampling: Restoration of original function(2)
Each sample serves as a weighting for a 2D sinc function.
),(h),(g yxyx
n m
)mY(Y
1sinc)nX(
X
1sinc)mY,nX(g yx
Nyquist/Shannon Theory: We must sample at twice the highest frequency in x and in y to reconstruct the original signal.
(No frequency components in original signal can be X21 or Y2
1 )
31-01-1387 7
Two Dimensional Sampling: Example
80 mm Field of View (FOV)
256 pixels
Sampling interval = 80/256 = .3125 mm/pixel
Sampling rate = 1/sampling interval = 3.2 cycles/mm or pixels/mm
Unaliased for ± 1.6 cycles/mm or line pairs/mm
31-01-1387 831-01-1387 8
FT of Shah function By similarity theorem
)()()]([
nssIII
t
xIIIFT
Example in spatial and frequency domain
)()(
nxx
III In time domain
Sampling process is Multiplication of infinite train of impulses III(x/∆x) with f(x)or convolution of III(u∆x ) with F(s)→ Replication of F(s)
31-01-1387 931-01-1387 9
Example in Time or Spatial domain
31-01-1387 1031-01-1387 10
Sampling theorem
A function sampled at uniform spacing can be recovered if
s2
1
Aliasing: = overlap of replicated spectra
31-01-1387 1131-01-1387 11
1) Truncation in Time Domain:Truncation of f(x) to finite duration T =Multiplying f(x) by Rect pulse of T = Convolving the spectrum with infinite sinx/x
Properties of Sampling I
T = N Δt (Truncation window) 1/T = 1/NΔt= Δs spectrum sample spacing (in DFT)
31-01-1387 1231-01-1387 12
Spectrum of truncation function is always infinite and Truncation destroy bond limitedness & produce alias.
This causes Unavoidable Aliasing
Since Truncation is:
Multiply f(t) with window
or convolve F(s) with narrow sin(x)/x Therefore, it extends frequency range (to infinite)
)(T
t
31-01-1387 1331-01-1387 13
2) There is a Sampling Aperture over which the signal is averaged at each sample point before applying Shah function
By convolve f(t) with aperture
or multiply F(s) with
This reduces high frequency of signal
)(1
t
s
ssin
Properties of sampling II
31-01-1387 1431-01-1387 14
)()()( sFsIIIsG
3) Since Sampling is multiplication of shah function with continues function Or convolution of F(s) with
Convolution of function with an impulse = copy of that function
Replicate F(s) every
1
Properties of sampling III
31-01-1387 1531-01-1387 15
4) Interpolation or Recovering original function (D/A)
ssssFs
ssG
1)()
2()( 1
1
Or convolving sampled g(x) with interpolation sinc
xs
xssxgFsFTxf
1
11
1
2
)2sin(2*)()()(
To recover original function, we should eliminate the replicas of F(s) and keep one.
Either Truncation in Freq should be done.
Properties of sampling IV
31-01-1387 1631-01-1387 16
31-01-1387 1731-01-1387 17
Review of Digitizing Parameters Depend on digitizing equipment:
Truncation window Max F.O.V of image
Sampling aperture Sensitivity of scanning spot
Sampling spacing Spot diameter (adjustable)
Interpolation function Displaying spot
31-01-1387 1831-01-1387 18
To have good spectra resolution (small Δs) and minimum aliasing, parameters N , T and Δt defined.
Review of Sampling Parameters
Δt as small as possible small ΔsT as long as possible compressed FT
To control aliasing: - Bigger sampling aperture
- Smaller sampling spacing (over same filter) - Adjust image freq. Sm at most Sm=1/2Δt
31-01-1387 1931-01-1387 19
Anti aliasing Filter:
1) Using rectangular aperture twice spacing Energy at frequency above S0>1/2Δt is attenuated. Original image freq. F(s) from 1/Δt reduce to 1/2Δt
31-01-1387 2031-01-1387 20
Anti aliasing Filter:
2) Using triangular aperture = 4 time of spacing Dies of frequency above 1/2Δt
31-01-1387 2131-01-1387 21
Examples of whole Sampling Process on a Band limited Signal
Original signal:
31-01-1387 2231-01-1387 22
Truncating the signal:
31-01-1387 2331-01-1387 23
Convolving signal with sampling aperture:
31-01-1387 2431-01-1387 24
Sampling the signal:
31-01-1387 2531-01-1387 25
Interpolating the sample signal (to recover original)
)(1
*)1
()1
(1
*]1
)([)( t
t
ttIII
tII
TIItfth
2])sin(
[)(*)sin(
])sin(
*)([)(
st
sttstIII
sT
sT
sT
sTTsFsI
31-01-1387 26
Discrete Fourier Transform
g(x) is a function of value for x
We can only examine g(x) over a limited time frame, L0 x
Assume the spectrum of g(x) is approximately bandlimited; no frequencies > B Hz.
Note: this is an approximation; a function can not be both time-limited and bandlimited.
f(x)
x
L2B
1 L
1-B -B
F(u)
31-01-1387 27
Sampling and Frequency Resolution
We will sample at 2B samples/second to meet the Nyquist rate.
f(x)
x
L2B
1 L
1-B -B
F(u)
BL2L
NB2
1 We sample N points.
resolutionfrequency L
1
N
B2
samples of #
rangefrequency
31-01-1387 28
Transform of the Sampled Function
Another expression for comes from
)nB2(F)(F
-δf)(f
n
B2n
1-N
0nB2n
2B1
uu
xx
1N
0
2
1N
0
B2f(n))(F
B2-x
2Bf
B2
1)(F
n
i
n
nuu
nnu
e
F
)(f xF)(F u
u is still continuous.
Bn
B 221 ff(n) where
Views input as linearcombination of deltafunctions
31-01-1387 29
Transform of the Sampled Function (2)
To be computationally feasible, we can calculate at only a finite set of points.
Since f(x) is limited to interval 0<x<L , can be sampled every Hz.
1N
0
2B2f(n))(F
n
i nuu e
Bn
B 221 ff(n) where
)(F u
L 1)(F u
31-01-1387 30
Discrete Fourier Transform
Discrete Fourier Transform (DFT):
N2BL number of samples
Number of samples in x domain = number of samples in frequency domain.
BL
nmi
nLm en 2
21N
0
)(fF(m)F
N
nmi
n
BL
nmi
n
enen 21N
0
221N
0
)(f)(fF(m)
31-01-1387 31
Periodicity of the Discrete Fourier Transform
F(m) repeats periodically with period N
1) Sampling a continuous function in the frequency domain [F(u) ->f(n)]causes replication of f(n) ( example coming in homework)
2) By convention, the DFT computes values for m= 0 to N-1
DFT:
1-N to1
1 to0
0 m
2N
2N
DC component
positive frequencies
negative frequencies
N
nmi
n
BL
nmi
n
enen 21N
0
221N
0
)(f)(fF(m)
31-01-1387 32
Properties of the Discrete Fourier Transform
Nkmi
eknfD.F.T.
2
m)(F)(
1. Linearity
2. Shifting
Let f(n) F(m)
Example : if k=1 there is a 2 shift as m varies from 0 to N-1
bG(u)aF(u)bg(x)af(x)
If f(x) F(u) and g(x) G(u)
31-01-1387 33
If f(n) F(m)
Inverse Discrete Fourier Transform
)(fF(m)F(m)1N
0
2
N11 N neD.F.T.
m
i nm
Why the 1/N ? Let’s take a look at an example
f(n)= {1 0 0 0 0 0 0 0}
1
00000001
f)(F1N
0
2
n8
n
i nmem
for all values of m continued...
N = 8 = number of samples.
31-01-1387 34
Now inverse,
Inverse Discrete Fourier Transform (continued)
2N N, 0, mfor 0)(f
1
1)(f
1
11
mF)(f
18)0(f
F(m))(f
F(m))(f
N
NN
N
8
N
2
2
N1
1
0
1N
0
2
N1
81
1N
0
2
81
1N
0
2
N1
n
e
en
r
r
r
en
en
en
m
m
nm
nm
nm
i
i
NN
n
i
m
i
m
i
For n= 0,kernel is simple
For other values of n, this identitywill help
31-01-1387 3531-01-1387 35