Scribd
Upload
40 views31 pages

3-Spatial Frequency and Transform

This document discusses key concepts in computer vision including frequency, Fourier transforms, filtering, sampling, and aliasing. [1] Frequency is measured in cycles or periods per second, with higher frequencies corresponding to more rapid changes in signals like images. Fourier transforms represent how a signal is composed of different frequencies. [2] Filters like lowpass, highpass, and bandreject can be used to remove certain frequency components. Lowpass filters retain slow variations while removing high frequencies like edges and noise. [3] Sampling converts continuous signals to discrete values, which can cause aliasing if the sampling rate is too low. Aliasing occurs when high frequencies fold back and are interpreted as lower frequencies. Gaussian

Uploaded by

Basudha Pal
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
40 views31 pages

3-Spatial Frequency and Transform

This document discusses key concepts in computer vision including frequency, Fourier transforms, filtering, sampling, and aliasing. [1] Frequency is measured in cycles or periods per second, with higher frequencies corresponding to more rapid changes in signals like images. Fourier transforms represent how a signal is composed of different frequencies. [2] Filters like lowpass, highpass, and bandreject can be used to remove certain frequency components. Lowpass filters retain slow variations while removing high frequencies like edges and noise. [3] Sampling converts continuous signals to discrete values, which can cause aliasing if the sampling rate is too low. Aliasing occurs when high frequencies fold back and are interpreted as lower frequencies. Gaussian

Uploaded by

Basudha Pal
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 31

Motion & Geometry based methods

in Computer Vision
ECE 4078- MGCV
Concept of Frequency
• Frequency is measured in Hertz, cycles or periods per
second
– Higher frequency means more cycle per second
– Higher frequency means shorter period
– Zero frequency means infinite period, which means constant
signal
• Fourier transform of a function of time has units, cycles
per second
– FT of a function of pixels (eg an image) will have units cycles per
pixel
– Rapid changes in a signal corresponds to higher frequencies,
slow changes are represented by lower frequencies
– Rapid changes in the intensity in an image are high frequency
Concept of Frequency
• Slowest varying frequency corresponds to the average gray level of
an image
• Low frequency corresponds to the slowly varying component of an
image
– Located near the origin of FT of image, F (0,0)
– Eg: Smooth gray level variation of wall, floor, sky, etc.
• High frequency corresponds to faster and faster gray level changes
– Located farther away from origin
– Eg: Edges of objects, noise
Fourier Transform

• Fourier transform is linear

• Inverse Fourier Transform


– To recover the signal from its Fourier Transform

• Phase and magnitude:


– Consists of real and complex component
F(g(x,y)) (u,v) = FR(g) + i* FI(g)

– Difficult to draw complex function of the plane=> separately plot


magnitude (magnitude spectrum) and phase (phase spectrum)
X
Fourier Transform

• FT of a function at (u,v) depends on WHOLE function


– What is F(0,0)?
– Notice that the domain of integral is over whole domain of
function

• Hence,
– A local change in the function (e.g: set to zero a block of point) is
going to modify every point in Fourier domain
• Difficult to use as a representation : Cannot tell whether a pattern is
present just by looking at FT

• STFT and Wavelet transform


Lowpass Filter
• Ideal LPF: Passes without attenuation all frequencies within a circle
of radius from the origin, and “cuts off” all frequencies outside this
circle

• where D0 is a positive constant, and


• D(u,v) is the distance between a point (u,v) in the frequency domain
and the center of the P ×Q frequency rectangle
Lowpass Filter

(a) Original image of size 688 × 688 pixels. (b)–(c) Results of filtering using
ILPFs with cutoff frequencies set at radii values 10 and 30 respectively
Lowpass Filter
• Gaussian lowpass filter:
Highpass Filter
• Subtracting a LPF transfer function from 1 yields the
corresponding HPF filter transfer function

• Ideal HPF

• Gaussian HPF
Highpass Filter
Highpass Filter

• Image filtered by ideal HPF and Gaussian HPF

• (a) Smudged thumbprint. (b) Result of highpass filtering (a). (c) Result of
thresholding
Bandreject Filter

• Ideal and Gaussian bandreject filters


• If additive periodic noise can be approximated as 2D sin function
– Easy to locate the noise in the Fourier transform of the image
Bandreject Filter

• In above image, noise components lie on an approx. circle, Hence


use circularly symmetric bandreject filter
– To remove ONLY noise, make the width of filter smaller
– The above result CAN NOT be obtained using spatial filtering !!
Sampling and aliasing
• What information is lost when we convert continuous time signal to
digital?
– Fourier transform provide an insight

• Eg: gray circle represents the samples. Which is a good sampling?


Sampling
• Sampling:
– Passing from a continuous function to a collection of values on a
discrete grid
– Eg: Passing from irradiation at the back of a camera system to
the pixel value reported by camera.

• Sampling in one dimension


– Takes a 1D function and returns a vector whose elements are
values of that function at the sample point
• For our purpose, assume sample points to be integer
• ith component of f is f(xi)
Sampling
• Sampling in two dimension
– Similar to sampling in 1D
– Assume samples are drawn at points with integer coordinates=>
yield uniform rectangular grid (good model for most camera)
– Therefore, sampled images are rectangular arrays of finite size

– NOTE: Sampling can occur on nonregular grid (eg human retina)


Sampling
• Sampling in two dimension
– Formally, sample a 2D function, which gives an array

– ith, jth element of array

– Note: Samples are not always evenly spaced in


practical systems
• Eg: TV screen aspect ration is 4: 3 (width: height)
• Cameras accommodate this effect by spacing samples points
slightly farther apart horizontally than vertical=> no square
pixels
Aliasing
• Sampling result in loss of information
• Sampled signal = product of original signal with bed of
nails function (2D 𝛿 function)
– Using convolution theorem, FT of this product = product of
convolution of FT of these two function
– Recall: Convolving a function with a shifted 𝛿 function merely shift
the function

– => Hence, FT of sampled signal is sum of collection


of shifted version of the Fourier transform of signal

– FT(sample2D(f(x,y)) = σ∞ ∞
𝑖=−∞ σ𝑗=−∞ 𝐹(𝑢 − 𝑖, 𝑣 − 𝑗)
• Where FT(f(x,y)) = F(u,v)
Aliasing

• Suppose the shifted


version of FT(f(x,y))
does not intersect=>
can reconstruct the
signal from the sampled
version
Aliasing

• But if support regions


Do overlap =>can not
Reconstruct the signal
(as different copies of FT
will add in overlap region)

• A sampling rate equal


to twice the highest
frequency is called the
Nyquist rate.
• Sampling must exceed
the Nyquist rate.
Aliasing
• Here sampled version of an image : resampling by factors of two
(all scaled to same size)
– Note aliasing: high frequency alias down to low spatial freq
– Smallest image is an extremely poor representation of large image
Aliasing
• Magnitude of Fourier Transform
– F(0,0) at center
– Note: FT of resampled image is obtained by scaling the FT of original image and
then tiling the plane
– Interference between copies of original FT => cannot recover original signal
Aliasing

1/2 1/4 (2x zoom) 1/8 (4x zoom)

Source: S. Seitz/Noah Snavely


Smoothing and resampling
• How to reduce the content of the original FT outside the range |u|
<1/2, |v| <1/2?
• Use Gaussian
• Recall: FT of a Gaussian is a Gaussian and Gaussian die away fairly
quickly
• Thus, convolve the image with Gaussian or multiply the FT by Gaussian

G 1/8
G 1/4
Gaussian 1/2
Subsampling with Gaussian prefiltering

Gaussian 1/2 G 1/4 G 1/8

Source: S. Seitz/Noah Snavely


Compare with…..

1/2 1/4 (2x zoom) 1/8 (4x zoom)

Source: S. Seitz/Noah Snavely


• filter the image,
then subsample

F0 F1 F2

blur subsample blur subsample …


F0* H F1* H

Source: S. Seitz/Noah Snavely


Gaussian Pyramid
Low resolution

blur
G2  (G1 * gaussian)  2

G1  (G0 * gaussian)  2
blur

G0  Image
blur

High resolution

Source: S. Seitz/Noah Snavely


Gaussian Pyramid

Source: S. Seitz/Noah Snavely

You might also like