Shanghai Jiao Tong University

Instructor: Xu Zhao

Computer Vision Class No.: AU7005

Spring 2023

Xu Zhao @ Shanghai Jiao Tong university

Lecture 5: Image Filtering

 Contents

❖ Spatial ltering
๏ Linear filter

๏ Non-linear filter

❖ Frequency perspective
๏ Spatial frequency

๏ Fourier transform
fi

 Image formation

Camera Sensor Output Image

 Image filtering, why and how?

❖ Motivation
❖ Enhance an image (denoise, resize, etc)
❖ Extract information (texture, edges, etc)
❖ Detect patterns (template matching)
❖ Compute a function of the local neighborhood at each
pixel in the image
❖ Function specified by a “filter” or mask saying how to
combine values from neighbors.

 Linear filtering
I(m, n) h(m, n)

h[m, n] = å f [k , l ] I [m + k , n + l ]
k ,l

For a linear system, each output is a linear

combination of all the input values
Example - Box filter

f [× ,× ]
1 1 1

1 1 1

1 1 1

Slide credit: David Lowe (UBC)

 Example - Box filter
h[m, n] = å f [k , l ] I [m + k , n + l ]
I [.,.] k ,l h[.,.]
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0
0
0 0 0 90 90 90 90 90 0 0
f [× ,× ]
0 0 0 90 90 90 90 90 0 0
1 1 1

0 0 0 90 90 90 90 90 0 0 1 1 1

1 1 1
0 0 0 90 0 90 90 90 0 0

0 0 0 90 90 90 90 90 0 0

0 0 0 0 0 0 0 0 0 0

0 0 90 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

m =1, n = 1
k,l = [-1,0,1]

 Example - Box filter

h[m, n] = å f [k , l ] I [m + k , n + l ]
I [.,.] k ,l h[.,.]
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0
10
0

0 0 0 90 90 90 90 90 0 0
f [× ,× ]
0 0 0 90 90 90 90 90 0 0
1 1 1

0 0 0 90 90 90 90 90 0 0 1 1 1

1 1 1
0 0 0 90 0 90 90 90 0 0

0 0 0 90 90 90 90 90 0 0

0 0 0 0 0 0 0 0 0 0

0 0 90 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

m =2, n = 1
k,l = [-1,0,1]

 Example - Box filter

h[m, n] = å f [k , l ] I [m + k , n + l ]
I [.,.] k ,l h[.,.]
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0
20
0 10

0 0 0 90 90 90 90 90 0 0
f [× ,× ]
0 0 0 90 90 90 90 90 0 0
1 1 1

0 0 0 90 90 90 90 90 0 0 1 1 1

1 1 1
0 0 0 90 0 90 90 90 0 0

0 0 0 90 90 90 90 90 0 0

0 0 0 0 0 0 0 0 0 0

0 0 90 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

m =3, n = 1
k,l = [-1,0,1]

 Example - Box filter

h[m, n] = å f [k , l ] I [m + k , n + l ]
I [.,.] k ,l h[.,.]
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0
30
0 10 20

0 0 0 90 90 90 90 90 0 0
f [× ,× ]
0 0 0 90 90 90 90 90 0 0
1 1 1

0 0 0 90 90 90 90 90 0 0 1 1 1

1 1 1
0 0 0 90 0 90 90 90 0 0

0 0 0 90 90 90 90 90 0 0

0 0 0 0 0 0 0 0 0 0

0 0 90 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

m =4, n = 1
k,l = [-1,0,1]

 Example - Box filter

h[m, n] = å f [k , l ] I [m + k , n + l ]
I [.,.] k ,l h[.,.]
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0
30
0 10 20 30

0 0 0 90 90 90 90 90 0 0
f [× ,× ]
0 0 0 90 90 90 90 90 0 0
1 1 1

0 0 0 90 90 90 90 90 0 0 1 1 1

1 1 1
0 0 0 90 0 90 90 90 0 0

0 0 0 90 90 90 90 90 0 0

0 0 0 0 0 0 0 0 0 0

0 0 90 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

m =5, n = 1
k,l = [-1,0,1]

 Example - Box filter

h[m, n] = å f [k , l ] I [m + k , n + l ]
I [.,.] k ,l h[.,.]
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0
30
0 10 20 30

0 0 0 90 90 90 90 90 0 0
f [× ,× ]
0 0 0 90 90 90 90 90 0 0
1 1 1

0 0 0 90 90 90 90 90 0 0 1 1 1

1 1 1
0 0 0 90 0 90 90 90 0 0

？
0 0 0 90 90 90 90 90 0 0

0 0 0 0 0 0 0 0 0 0

0 0 90 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

m =4, n = 6
k,l = [-1,0,1]

 Example - Box filter

h[m, n] = å f [k , l ] I [m + k , n + l ]
I [.,.] k ,l h[.,.]
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0 10 20 30 30

0 0 0 90 90 90 90 90 0 0
f [× ,× ]
0 0 0 90 90 90 90 90 0 0
1 1 1

1 1 1
？
0 0 0 90 90 90 90 90 0 0
1 1 1
0 0 0 90 0 90 90 90 0 0

0 0 0 90 90 90 90 90 0 0 50

0 0 0 0 0 0 0 0 0 0

0 0 90 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

m =6, n = 4
k,l = [-1,0,1]

 Example - Box filter

h[m, n] = å f [k , l ] I [m + k , n + l ]
I [.,.] k ,l h[.,.]
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0 10 20 30 30 30 20 10

0 0 0 90 90 90 90 90 0 0
f [× ,× ] 0 20 40 60 60 60 40 20

0 0 0 90 90 90 90 90 0 0 0 30 60 90 90 90 60 30
1 1 1

0 0 0 90 90 90 90 90 0 0 1 1 1 0 30 50 80 80 90 60 30
1 1 1
0 0 0 90 0 90 90 90 0 0 0 30 50 80 80 90 60 30

0 0 0 90 90 90 90 90 0 0 0 20 30 50 50 60 40 20

0 0 0 0 0 0 0 0 0 0 10 20 30 30 30 30 20 10

0 0 90 0 0 0 0 0 0 0 10 10 10 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0
 Box filter

❖ What does it do? f [× ,× ]

❖ Replaces each pixel with an
1 1 1
average of its neighborhood
1 1 1
❖ Achieve smoothing effect
(remove sharp features) 1 1 1

Smoothing with box filter

f [× ,× ]
1 1 1

1 1 1

1 1 1
 Integral image
❖ Don’t forget to use integral image to improve computational
efficiency for utilization of bunch of box filters.
3.2 Linear filtering 121

3 2 7 2 3 3 5 12 14 17 3 5 12 14 17
1 5 1 3 4 4 11 19 24 31 4 11 19 24 31
5 1 3 5 1 9 17 28 38 46 9 17 28 38 46
4 3 2 1 6 13 24 37 48 62 13 24 37 48 62
2 4 1 4 8 15 30 44 59 81 15 30 44 59 81

(a) S = 24 (b) s = 28 (c) S = 24

Figure 3.17 Summed area tables: (a) original image; (b) summed area table; (c) computation
of area sum. Each value in the summed area table s(i, j) (red) is computed recursively from
 Correlation and Convolution

❖ 2D correlation: ❖ Convolution is the same as

correlation with a 180°
h[m, n] = å f [k , l ] I [m + k , n + l ] rotated lter kernel.
k ,l
❖ Correlation and
❖ 2D convolution: convolution are identical
when the lter kernel is
h[m, n] = å f [k , l ] I [m - k , n - l ] symmetric.
k ,l
fi
fi

 Key properties of linear filters

❖ Linearity:

imfilter(I, f1+f2)=imfilter(I,f1)+imfilter(I,f2)

❖ Shift invariance: same behavior given intensities

regardless of pixel location m,n

imfilter(I,shift(f)) = shift(imfilter(I,f))

❖ Any linear, shift-invariant operator can be represented as a

convolution.

 Convolution properties
❖ Commutative: a * b = b * a
❖ Conceptually no difference between filter and signal
❖ Associative: a * (b * c) = (a * b) * c
❖ Often apply several filters one after another: (((a * b1) * b2) * b3)
❖ This is equivalent to applying one filter: a * (b1 * b2 * b3)
❖ Correlation is _not_ associative (rotation effect)
❖ Distributes over addition: a * (b + c) = (a * b) + (a * c)
❖ Scalars factor out: ka * b = a * kb = k (a * b)
❖ Identity: unit impulse e = [0, 0, 1, 0, 0], a * e = a

Source: S. Lazebnik

Correlation for template matching

Let’s see if we can use correlation to ‘ nd’ the parts of the image
that look like the lter.

D I = correlate2d( D, f, ‘same’ )

Correct
location
+

Response peak

f It does’t work. Why?

[Thanks to Robert Collins @ Penn State]

fi
fi
Correlation for template matching
D
❖ It does’t work. Why?

h[m, n] = å f [k , l ] I [m + k , n + l ]
k ,l

❖ As brightness in I increases,
the response in h will
increase, as long as f is
positive.
f
❖ Overall brighter regions will
give higher correlation
response.

Correlation for template matching

❖ Normalization correlation. Subtract the mean so zero centered.
Score is higher only when dark parts match and when light parts
match.
❖ This is a great way to build simple, fast pattern detectors that require
little computation.

I2 = correlate2d( D2, f2, ‘same’ )

D2 = D – np.mean(D) f2 = f – np.mean(f)
Correlation for template matching

❖ What happens with convolution?

I2 = convolve2d ( D2, f2, ‘same’ )

D2 = D – np.mean(D) f2 = f – np.mean(f)
Correlation for template matching

Forsyth and Ponce

 Gaussian filters
x

0.003 0.013 0.022 0.013 0.003

0.013 0.059 0.097 0.059 0.013
y 0.022 0.097 0.159 0.097 0.022
0.013 0.059 0.097 0.059 0.013
0.003 0.013 0.022 0.013 0.003

x y 5 x 5, s = 1

Weight contributions of neighboring pixels by nearness

Smoothing with Gaussian filter
Smoothing with box filter
 Gaussian filters
❖ Remove “high-frequency” components from the image (low-pass filter)
❖ Images become more smooth
❖ Gaussian convolved with Gaussian is another Gaussian
❖ So can smooth with small-width kernel, repeat, and get same result
as larger-width kernel would have
❖ Convolving twice with Gaussian kernel of width σ is same as
convolving once with kernel of width  σ√2
❖ Separable kernel
❖ Factors into product of two 1D Gaussians
Source: K. Grauman
 Gaussian filters
❖ What parameters matter here?
❖ Size of kernel or mask

σ = 5 with 10 x 10 kernel σ = 5 with 30 x 30 kernel

 Gaussian filters
❖ What parameters matter here?
❖ Variance of Gaussian: determines extent of smoothing

σ = 5 with 10 x 10 kernel σ = 5 with 30 x 30 kernel

Source: K. Grauman
Gaussian filters: Separability

Source: D. Lowe
 Separability
2D convolution
(center location only)

The lter factors

into a product of 1D
lters:

Perform convolution
* =
along rows, followed by
convolution
along the remaining *
=

column:
Source: K. Grauman
fi
fi











 Separability

❖ Why is separability useful in practice?

❖ MxN image, PxQ filter
❖ 2D convolution: ~MNPQ      multiply-adds
❖ Separable 2D:    ~MN(P+Q) multiply-adds
❖ Speed up = PQ/(P+Q)
❖ 9x9 filter = ~4.5x faster

 Practical matters
❖ What about near the edge?
❖ The filter window falls off the edge of the image
❖ Need to extrapolate
❖ Methods
❖ clip filter (black)
❖ wrap around
❖ copy edge
❖ reflect across edge
Source: S. Marschner

Convolution in Convolutional Neural Networks

❖ Convolution is the basic operation in CNNs

❖ Learning convolution kernels allows us to learn which
`features’ provide useful information in images
 Non-linear filters
❖ Median filter
❖ Operates over a window
by selecting the median
intensity in the window.
❖ ‘Rank’ filter as based on
ordering of gray levels
❖ E.G., min, max, range
filters

 Non-linear filters
❖ Median filter
❖ Operates over a window by selecting the median
intensity in the window.
❖ ‘Rank’ filter as based on ordering of gray levels
❖ E.G., min, max, range filter
❖ Very efficient to remove salt-and–pepper noise
❖ Variants: α-trimmed mean, weighted median

Salt and
pepper noise Median
filtered

Plots of a row of the image

Source: M. Hebert
 Non-linear filters
❖ Median filter
❖ Median filter is edge preserving

Source: K. Grauman

Frequency perspective

What can you see?

Salvador Dali (萨尔瓦多·达利),1976

 Hybrid Images

Aude Oliva & Antonio Torralba & Philippe G Schyns, SIGGRAPH 2006
 Hybrid Images

Aude Oliva & Antonio Torralba & Philippe G Schyns, SIGGRAPH 2006
 Hybrid Images

Aude Oliva & Antonio Torralba & Philippe G Schyns, SIGGRAPH 2006
Aude Oliva & Antonio Torralba & Philippe G Schyns, SIGGRAPH 2006
 Sampling
❖ Why does a lower resolution image still make sense to us? 
What do we lose?

Source: James Tompkin

 Subsampling

❖ Subsampling by a factor of 2

Throw away every other row

and column to create a 1/2 size
image Source: James Tompkin
Figure from David Forsyth
1/2

1/4 2x subsample 1/8 4x subsample

Steve Seitz
Sampling and aliasing
 Sampling and aliasing
1D example (sinewave):

Source: S. Marschner
 Sampling and aliasing
1D example (sinewave):

Source: S. Marschner
 Aliasing problem
❖ Sub-sampling may be dangerous….
❖ Characteristic errors may appear:
❖ “car wheels rolling the wrong
way in movies
❖ “checkerboards disintegrate in
ray tracing”
❖ “striped shirts look funny on
color television”
❖ Moiré patterns

Source: James Tompkin

 Aliasing and Moiré patterns

❖ Moire pattern, known as the moire effect
❖ Useful in metrology, strain analysis or even document authentication and anti-
counterfeiting. In other situations moire patterns may be unwanted.
❖ This visual perception occurs when a fine pattern on your subject meshes with
the pattern on the imaging chip of your camera. Difference in frequency.

Aliasing in video

Slide by Steve Seitz

Nyquist-Shannon Sampling Theorem
❖ When sampling a signal at discrete intervals, the sampling frequency must
be ≥ 2 × fmax
❖ fmax = max frequency of the input signal
❖ This will allows to reconstruct the original perfectly from the sampled
version
❖

v v v good

bad
 How to fix aliasing?

❖ Sample more often (better sensor)

❖ Get rid of all frequencies that are greater than half the
new sampling frequency
❖ Will lose information
❖ But it’s better than aliasing
❖ Apply a smoothing (low pass) filter

Hays

Anti-aliasing

Forsyth and Ponce 2002

Algorithm for downsampling by factor of 2

❖ Start with image(h, w)

❖ Apply low-pass filter:

im_blur = imfilter( image, fspecial(‘gaussian’, 7, 1) )

❖ Sample every other pixel

im_small = im_blur( 1:2:end, 1:2:end )

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

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

Subsampling with Gaussian pre- ltering

fi
Human spatial frequency sensitivity
Why do we get different, distance-dependent
interpretations of hybrid images?

Hays
 Clues from Human Perception
Early processing in humans lters for orientations and scales of frequency.

Early Visual Processing: Multi-scale edge and blob lters

Source: James Tompkin
fi
fi
Campbell-Robson contrast sensitivity curve v

Perceptual cues in the mid-high frequencies dominate perception.

Contrast decrease (log)

Frequency increase (log)

Source: James Tompkin
 Salvador Dali
“Gala Contemplating the Mediterranean Sea,
which at 30 meters becomes the portrait
of Abraham Lincoln”, 1976

 Fourier Transforms

❖ The tool manipulating frequency. Sampling, De-noise,

filtering…
❖ Linear, space invariant operations are just diagonal
operations in the frequency domain
❖ Linear convolution is multiplication in the frequency
domain

 Fourier series
❖ A bold idea (1807):
❖ Any univariate function can
be rewritten as a weighted
sum of sines and cosines of
different frequencies.
"

! "! cos('() + +! sin('()

!#$

❖ Add enough of them to

get any signal g(t) you
want!
Jean Baptiste Joseph Fourier (1768-1830)

 t = [0,2], f = 1

g(t) = (1)sin(2πf t) + (1/3)sin(2π(3f) t)

g = +

t Coefficient

Slides: Efros
Wikipedia-Fourier transform
 Sine/cosine and circle

Wikipedia - Unit circle

Square wave (approx.)

Mehmet E. Yavuz
Square wave (approx.)

Mehmet E. Yavuz
One series in each of x and y

Generative Art, じゃがりきん, Video, 2018

 Amplitude-phase form
❖ Add phase term to shift
cos values into sin values

"
! "! sin('( + ∅! )
!#$

Phase

[Peppergrower; Wikipedia]



 Amplitude-phase form
❖ Add component of infinite frequency
‣ = mean of signal over period

‣ = value around which signal fluctuates

# !!
!!
+ $ !" sin()* + ∅" ) 2
2
"$%

Average of signal
over period



Amplitude-phase form

[Peppergrower; Wikipedia]

 Fourier Transform
❖ Stores the amplitude and phase at each frequency:
❖ For mathematical convenience, this is often
notated in terms of real and complex numbers
❖ Related by Euler’s formula
❖ Amplitude encodes how much signal there is at a
particular frequency

A = ± Re(j ) 2 + Im(j ) 2

❖ Phase encodes spatial information (indirectly)

Im(j )
f = tan-1

Re(j )

 Fourier analysis in images

Spatial domain images

Fourier decomposition amplitude images

http://sharp.bu.edu/~slehar/fourier/fourier.html#filtering More: http://www.cs.unm.edu/~brayer/vision/fourier.html
 Signals can be composed
Spatial domain images

+ =

Fourier decomposition amplitude images

http://sharp.bu.edu/~slehar/fourier/fourier.html#filtering More: http://www.cs.unm.edu/~brayer/vision/fourier.html
 Natural images
Natural image Fourier decomposition amplitude image

① What does it mean to be at pixel x,y?

② What does it mean to be more or less bright in the Fourier decomposition
image?
Source: James Tompkin

 Amplitude / Phase

❖ Amplitude tells you “how

much”
❖ Phase tells you “where”
❖ Translate the image?
❖ Amplitude unchanged
❖ Adds a constant to the
phase.
[Peppergrower; Wikipedia]

Amplitude / Phase

Efros
 Amplitude / Phase
Amplitude Phase

Efros
Amplitude / Phase

Efros
 Amplitude / Phase

Amplitude Phase

Efros
 Amplitude / Phase
Cheebra
Zebra phase, cheetah amplitude Cheetah phase, zebra amplitude

Efros
 Amplitude / Phase
❖ The frequency amplitude of natural images are quite
similar
❖ Heavy in low frequencies, falling off in high frequencies
❖ Will any image be like that, or is it a property of the
world we live in?
❖ Most information in the image is carried in the phase, not
the amplitude
❖ Not quite clear why
Efros
Properties of Fourier Transforms

❖ Linearity
F[ax(t ) + by (t )] = a F[ x(t )] + b F[ y (t )]

❖ Fourier transform of a real signal is symmetric about the

origin
❖ The energy of the signal is the same as the energy of its
Fourier transform

 The Convolution Theorem

❖ The Fourier transform of the convolution of two

functions is the product of their Fourier transforms
F[ g * h] = F[ g ] F[ h]

❖ Convolution in spatial domain is equivalent to

multiplication in frequency domain
-1
g * h = F [F[ g ] F[h]]

Filtering in spacial domain

1 0 -1
2 0 -2
1 0 -1

* =
Convolution

Hays
Filtering in frequency domain

Element-wise
product

Fourier Fourier
transform transform
Inverse
Fourier transform

Slide: Hoiem
Frequency manipulation
Low and high pass filtering
Removing frequency bands

Brayer
Why does the Gaussian filter give a nice smooth image, but
the square filter give edgy artifacts?

Gaussian Box filter

Why do we have those lines in the image?

❖ Sharp edges in the image need all frequencies to

represent them.

¥
1
= Aå sin(2p kt )
k =1 k
 Box filter / sinc filter duality
❖ What is the spatial representation of the hard cutoff (box) in the
frequency domain?
❖ http://madebyevan.com/dft/
Box filter Sinc filter sinc(x) = sin(x) / x

Spatial Domain Frequency Domain

Frequency Domain Spatial Domain
 Gaussian filter duality
❖ Fourier transform of one Box filter (spatial)
Frequency domain
magnitude
Gaussian is another Gaussian
(with inverse variance).
❖ Why is this useful?
❖ Smooth degradation in
frequency components Gaussian filter Frequency domain
(spatial) magnitude

❖ No sharp cut-off
❖ No negative values
❖ Never zero (infinite extent)
Source: James Tompkin
 Ringing artifacts -> ‘Gibbs effect’
Where infinite series can never be reached
Gaussian Box filter

Source: James Tompkin

 Is convolution invertible?

❖ If convolution is just multiplication in the Fourier

domain, isn’t deconvolution just division?
❖ Sometimes, it clearly is invertible (e.g. a convolution
with an identity filter)
❖ In one case, it clearly isn’t invertible (e.g. convolution
with an all zero filter)
❖ What about for common filters like a Gaussian?

 Convolution

* =

FFT FFT iFFT

.x =

Hays
 Deconvolution

iFFT FFT FFT

= ./

Hays
But under more realistic conditions
Random noise, .000001 magnitude

iFFT FFT FFT

= ./

Hays
But under more realistic conditions
Random noise, .0001 magnitude

iFFT FFT FFT

= ./

Hays
But under more realistic conditions
Random noise, .001 magnitude

iFFT FFT FFT

= ./

Hays
 Deconvolution is hard.

❖ Active research area.

❖ Even if you know the filter (non-blind deconvolution), it is still hard
and requires strong regularization to counteract noise.
❖ If you don’t know the filter (blind deconvolution), then it is harder
still.
Earl F. Glynn
Earl F. Glynn