Professional Documents
Culture Documents
Yoram Tal
Image Transforms
• Fourier, FFT
• Wavelets
• Radon, Hough
75
Forward Transform
Source Image X
direct indirect
processing processing
Backward Transform
Output Image Y
76
38
Dr. Yoram Tal
Image Frequency
domain domain
n=N
Every digital signal
f ( x) = ∑F e
n=− N
n
inπx / L
may be expanded in a
sum of sine & cosine
Euler: eiθ = cos(θ) + i sin(θ) functions 77
78
39
Dr. Yoram Tal
79
80
40
Fourier Transform
Z1
i!t
X(!) = x(t)e
1
Inverse
Z1
i!t
x(t)= 21 X(!)e d!
1
N
X1
b =
X xn e
2 in
N = 0; 1; :::N 1
n=0
Inverse
N
X1
xn = N1 b e2
X
in
N n = 0; 1; :::N 1
=0
1
Properties
Separability
N
X1 NX1 2 i(n +m )
F = N12 fnm e N ; = 0; 1; :::; N 1
n=0 m=0
N
X1 2 in
= N1 Fn e N
n=0
where
N
X1 2 im
Fn = N1 Fnm e N
m=0
Translation
2 in 0
F 0
$ f ne N
2 in 0
F e N $ fn n0
1
Convolution
Continuous
R
f g(x) = f (y)g(x y)dy
Discrete
N
X1
1
f g(x) = N f (n)g(x n)
n=0
Fourier
Correlation
Continuous
R
f g(x) = f (y)g(x + y)dy = f (y) g( y)
Discrete
N
X1
1
f g(x) = N f (n)g(x + n)
n=0
Fourier
1
Dr. Yoram Tal
y z { | | } z ~ ~ { } ~
f (x ) (t = 0)
u (x , t ) =
î G 2 t ⊗ f (x ) ( ) (t > 0)
where
x2
Gσ (x ) =
1
⋅ exp − f : R2 → R
2πσ 2
2σ
2
99
¡ ¢ £ ¤ ¥ ¥ ¥ ¢ ¦ § ¡ ¨ ¦ © ª ¢ « ¬ ® ¨ « ¯ ° ± ² § ¡ ¨ ¦ © ª ¢ « ¬ ³ ¦ ± ª ´ µ ¡ ¶ · ¦ ° ¸
¹ º » ¼ ½ ¾ ¿ À Á Â Ã Ä Å Å Æ Ç È É É Ê Ë Ì Í Ë Î Ï Ð Ñ Ò Ó Ô Õ Ö Ñ Ò × Ø Ô Õ Ù Ð Ñ Ú Û
100
50
Heisenberg Uncertainty Principle (band limited in physical space
and Fourier space)
R +∞ √
Normalize f(x) so that −∞ |f (x)|2 dx = 1 and limx→∞ xf (x) = 0
Let F(k) be the Fourier transform
R +∞ of f(x).
Then by Parseval’s theorem −∞ |F (k)|2 dk = 1
sZ sZ
Z ∞ ∞ ∞
0
| x(f (x)f (x)dx| ≤ |xf (x)|2 dx |f 0 (x)|2 dx by the Schwarz inequal
−∞ −∞ −∞
sZ sZ
∞ ∞
= |xf (x)|2 dx |ikF (k)|2 dk by Parseval’s equality
−∞ −∞
Z ∞ Z ∞
1 1
x(f (x)f 0 (x)dx = x [f (x)]2 |∞
−∞ − [f (x)]2
−∞ 2 −∞ 2
1 1
=0− =− by integration by parts
2 2
Hence sZ sZ
∞ ∞
2
1
|xf (x)| dx |ikF (k)|2 dk ≥
−∞ −∞ 2