COMPUTER VISION

(CSE3010)
Module-1
CONVOLUTION &
LINEAR FILTERS Dr. Paras Jain
Associate Professor
School of Computing Science
Engineering and Artificial Intelligence
LINEAR FILTERS – INTRODUCTION TO DISCRETE TIME SIGNALS
 LINEAR FILTERS – IMPULSE
RESPONSE

Impulse response of a continuous system

Impulse response of a discrete system

Impulse function
1DIMENSIONA
L
CONVOLUTION
BRIEF LOOK
Convolution of two signals f and g
LINEAR FILTERS-TIME DOMAIN VS FREQ DOMAIN
IMPULSE, STEP
AND FREQUENCY
RESPONSE OF A
LINEAR FILTER

Every linear filter has an impulse response, a step response

and a frequency response
TWO-DIMENSIONAL
FOURIER
TRANSFORM
 EDGE EFFECTS IN DISCRETE
CONVOLUTIONS
 Fourier transform is a type of integral transform

The process takes a complex valued function of x, y and returns

a complex valued function of u, v (images are complex valued functions
with zero imaginary component)
 EDGE EFFECTS IN DISCRETE
CONVOLUTIONS

 These terms are sinusoids on the x, y plane, whose

orientation and frequency are given by u, v

 tan θ = v/u

 The gradient of this term is perpendicular to lines

where ux+vy is constant, and the frequency of the
sinusoid is 𝑢𝑢2 + 𝑣𝑣2

 These sinusoids are often referred to as spatial frequency

components
 EDGE EFFECTS IN DISCRETE
CONVOLUTIONS
 Inverse Fourier transform , IDFT for discrete
 It is useful to be able to recover a signal from its Fourier transform
LINEAR FILTERS -1D &
2D CONVOLUTION
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION

Continuous time
convolution

Discrete time
convolution
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION

Cross check with the

analytical method
LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION

Cross check with the analytical method

LINEAR FILTERS – INTRODUCTION TO 1D CONVOLUTION

Cross check with the analytical method

 LINEAR FILTERS - CONVOLUTION

 Convolution between two given signals, 𝒇 𝒔

and 𝒉𝒉 𝒔 is given by the following integral,
where
𝒚 denotes the convolution output:
∞
𝒚 =𝒇
𝒇 ∗𝒉
𝒉𝒙 � 𝒔𝒇 . 𝒉𝒉 𝒙 𝒅
= − 𝒔 𝒅𝒔
𝒔
−
∞
= �
∞
𝒉𝒉 𝒔 . 𝒇 𝒙 𝒅
− 𝒔 𝒅𝒔
𝒔
CONVOLUTION –
A CLOSER LOOK
 LINEAR FILTERS - CONVOLUTION

Convolution of a discrete- time input signal x[k]

and impulse response h[n- k]. (commutative)
 LINEAR FILTERS – FILTER KERNELS

The pattern of weights used

for a linear filter is usually
referred to as the kernel of
the filter
 LINEAR FILTERS – 2D
CONVOLUTION
 Convolution between a kernel (typically a 3x3 or 5x5 matrix)
and an image

Kernel of the
2D filter
Image
DIFFERENT FILTER KERNELS FOR 2D CONVOLUTION

Sobel Kernel

Gaussian Kernel

Box-blur Kernel
LINEAR FILTERS –
TYPES
LINEAR FILTERS
– TYPES
LINEAR FILTERS
– TYPES
LINEAR FILTERS
– TYPES
LINEAR FILTERS
– TYPES
SMOOTHING BY
AVERAGING
 SMOOTHING BY
 AVERAGING
Reduce the effects of noise by replacing each pixel with a
weighted average of its neighbours, a process often referred to as
blurrin
smoothing or
g

Original Image Smoothed Image

SMOOTHING BY
AVERAGING
SMOOTHING BY
AVERAGING
  Although a uniform local average may
seem to give a good blurring model, it
generates effects that are not usually
seen in defocussing a lens. The images
above compare the effects of a uniform
local average with weighted average

 The degree of blurring in each case is

about the same, but the uniform average
produces a set of narrow vertical and
horizontal bars — an effect often known
as ringing

 The bottom row shows the weights used to

blur the image, themselves rendered as an
image; bright points represent large
values and dark points represent small
values (in this example the smallest values
THE 2D GAUSSIAN
SMOOTHING
FILTER
 GAUSSIAN SMOOTHING
FILTER
 Case of weighted averaging
Coefficients are 2D Gaussian
Gives more weight at the central pixels and less
weights to the neighbors
 The farther away the neighbors, the smaller
the weight

1
−
𝑥𝑥 2 + 𝑦
𝐺𝐺𝜎𝜎(x, y) = 2 exp 𝑦 2

2
𝜋𝜋 2𝜋𝜋2
 GAUSSIAN SMOOTHING
FILTER

 The symmetric Gaussian kernel in

2D. This view shows a kernel scaled so
that its sum is equal to one

 Convolution with this kernel forms

a weighted average which stresses
the point at the center of the
convolution window, and
incorporates little contribution
from those at the boundary.
𝐺𝐺𝜎𝜎(x, y) = 1 2 exp − 𝑥𝑥2 +
2
𝜋𝜋
𝑦𝑦2 2𝜋𝜋
GAUSSIAN SMOOTHING
FILTER
BOX FILTER VS
GAUSSIAN
GAUSSIAN SMOOTHING – DIFFERENT
SCALES
SMOOTHING BY GAUSSIAN
FILTER
The standard deviation of
the gaussian determines
the amount of smoothing
Gaussian theoretically has
infinite support, but we
need a filter of finite size
For a 98.76% of the area,
we need +/ -2.5σ
 + / - 3σ covers over 99% of
the area
SHIFT-
INVARIANT
LINEAR SYSTEMS
 SHIFT INVARIANT LINEAR
SYSTEMS
 Most imaging systems have, to a good approximation, three
significant properties
 SHIFT INVARIANT LINEAR
SYSTEMS
 Most imaging systems have, to a good approximation,
three significant properties

In a shift invariant system, the response to a translated

stimulus is just a translation of the response to the
stimulus

The response of a shift invariant linear system to a

stimulus is obtained by convolution
SAMPLING AND
ALIASING
 SAMPLING AND
ALIASING
The Nyquist sampling theorem provides a
prescription for the nominal sampling interval
required to avoid aliasing
It may be stated simply as follows:
The sampling frequency should be at least twice the
highest frequency contained in the signal. Or in
mathematical terms: fs >= 2fc
𝑠 𝑓𝑠𝑓
≥where
2𝑓𝑓𝑐𝑐 fs is the sampling frequency
(how often samples are taken per unit of time or space),
and 𝑐𝑓𝑓𝑐is the highest frequency contained in
the signal.
SAMPLIN
G
 The sampling frequency
should be at least twice
the highest frequency
contained in the signal.
Or in mathematical
terms:
𝑠𝑓𝑠𝑓≥ 2𝑓
𝑓𝑐
𝑐
where fs is the sampling
frequency (how often
samples are
taken per unit
of time or
spa ce), and 𝑐𝑓𝑓𝑐is the highest
frequency contained in the
signal.
SAMPLIN
G
 The sampling frequency
should be at least twice
the highest frequency
contained in the signal.
Or in mathematical
terms:
𝑠𝑓𝑠𝑓≥ 2𝑓
𝑓𝑐
𝑐
where fs is the sampling
frequency (how often
samples are
taken per unit
of time or
spa ce), and 𝑐𝑓𝑓𝑐is the highest
frequency contained in the
signal.
SAMPLIN
G
 The sampling frequency
should be at least twice
the highest frequency
contained in the signal.
Or in mathematical
terms:
𝑠𝑓𝑠𝑓≥ 2𝑓
𝑓𝑐
𝑐
where fs is the sampling
frequency (how often
samples are
taken per unit
of time or
spa ce), and 𝑐𝑓𝑓𝑐is the highest
frequency contained in the
signal.
ALIASIN
G
 The sampling frequency
should be at least twice
the highest frequency
contained in the signal.
Or in mathematical
terms:
𝑠𝑓𝑠𝑓≥ 2𝑓
𝑓𝑐
𝑐
where fs is the sampling
frequency (how often
samples are
taken per unit
of time or
spa ce), and 𝑐𝑓𝑓𝑐is the highest
frequency contained in the
signal.
ALIASIN
G
 The sampling frequency
should be at least twice
the highest frequency
contained in the signal.
Or in mathematical
terms:
𝑠𝑓𝑠𝑓≥ 2𝑓
𝑓𝑐
𝑐
where fs is the sampling
frequency (how often
samples are
taken per unit
of time or
spa ce), and 𝑐𝑓𝑓𝑐is the highest
frequency contained in the
signal.