Digital Image Processing - COMP4173
Lecture 5: Fourier Transform
Prof. Hongjian Shi (时红建）
Department of Computer Science and Technology
BNU-HKBU United International University
Email:
[email protected]Office: T3-601-R3;
Office Hours: Tues., Wed., Thur. 9:00-11:50; Wed. 16:00-16:50
1
Transforms in Frequency Domain
• u, v are called the transform variables
• T(u, v) is called the forward transform of f(x, y)
• R operation is called the inverse transform
• s(x, y, u, v) is called the inverse transformation kernel
• s(x, y, u, v) is separable if s(x, y, u, v) = s1(x, u) s2(y, v)
Fourier Transform
• A continuous signal can be
represented as weighted sum of
sinusoids (basis functions)
Square Wave
Structure of Sinusoids
Fourier coefficients
Where Fourier transform comes from?
• Theorem: Any continuous function over −𝜋, 𝜋 can be expressed as a
sum of sinusoids (actually this is the inverse Fourier transform for a
continuous function with a period 2𝜋:
1 𝜋 −𝑛2𝜋𝜔𝑡
= 𝑓(𝑡)𝑒 dt
2𝜋 −𝜋
Fourier coefficients
• How are the 𝐶𝑛 computed? That is, the Fourier transform comes from.
1D Fourier Transform
f(t) is a non periodical function
Inverse Fourier Transform:
f(t) is a function with period T
DFT of a
sampled
function
1D DFT
|F(u)| - magnitude or
spectrum
R(u) – real part
I(u) – imaginary part
- Phase spectrum or angle between R(u) and I(u)
- Power spectrum or spectral density
1D Discrete Fourier Transform
a 𝑢, 𝑛 = 𝑒 −2𝜋𝑢𝑛/𝑁 are bases
𝑏 𝑢, 𝑛 = 𝑒 2𝜋𝑛𝑢/𝑁
1D DFT of Different Lengths
real(A) imag(A) real(A) imag(A)
1D DFT of Signals
• 1D Fourier transform pairs (spatial vs. frequency):
1D DFT of Signals – cont.
1D DFT of Signals – cont.
• Periodicity: F(u) has a period of M
• Symmetry: magnitude or spectrum centered at the origin
Aliasing
Sampling Theorem (Nyquist):
2D impulse:
…… there are similar reasoning for 2D pulse trains and the Fourier transform
Discrete Fourier Transform (DFT)
• Fourier transform:
• Inverse Fourier
Transform:
DFT and IDFT
• DFT does not contain all frequencies forming an image, only a set
of samples large enough to describe the spatial domain image
• The number of frequencies corresponds to the number of pixels
in spatial domain image and so the image size in spatial domain
is same as that in frequency domain
• For an image of size MxN, DFT and IDFT are:
From 1D DFT to 2D DFT
?
2D DFT
Domain: frequency – cycles/sec, or cycles/meter, or generally cycles/unit variable.
Range: complex plane
2D DFT – cont.
- Magnitude / Spectrum
- Phase spectrum
- Power spectrum
2D DFT – cont.
• Periodicity:
F(u,v) has a
period of N in
horizontal and
M in vertical
directions
• Symmetry: its
magnitude is
centered at the
origin
Return back to 8-10)
Aliasing due to undersampling
Why we need transforms in frequency
domain?
• Better image processing – enhancement, high contrast, more
insights, smooth, noise reduction etc.
• Fast computation – convolution instead of multiplication
• Inverse transform to recover images from transformed data
• Simple representation of images
• Efficient storage and transmission for images and videos
• Energy compaction
• Representative frequency basis
