SIGNAL PROCESSING

CONVOLUTION

Franck Dufrenois MISC

Convolution : applications

signal processing

s(t)
r(t)
d

Restauration of vinyls: removing of cracking

noises from soiled vinyl records Distance from the target plane: the signal r(t) is
noised and received with a delay with respect to
s(t)
Convolution: applications

image processing

Smoothing filter: low pass filter Enhancing filter : high pass filter Derivating filter
convolution: definition

It has a key role in continuous or discret signal. This operation is realized from a measuring tool
(microscope, camera, optical senses,..) which transforms or filters an input x(t) to an input y(t):

Convolution is defined by :

Continuous version
convolution: Impulse response

h(t) is named impulse response because it is the result of an input impulse signal 𝛿(𝑡)

The system is linear and invariant then the corresponding answer to

We have seen previously that:

The, we deduce that the overall response of the system

Intégrale de convolution
convolution: Graphical meaning
1 x(u) 1 x(u)
ℎ −0.25 − 𝑢 ℎ 0.5 − 𝑢

-0.5 -0.25 +0.5 -0.5 0 +0.5

+∞ +∞
න 𝑥 𝑢 ℎ −0.25 − 𝑢 𝑑𝑢 = 3/32 න 𝑥 𝑢 ℎ 0.5 − 𝑢 𝑑𝑢 = 1/8
−∞ −∞
x(u)
1 1 x(u)
ℎ 0−𝑢

-0.5 +0.5
-0.5 0 +0.5
h(u) h(-u) න
+∞
𝑥 𝑢 ℎ 0 − 𝑢 𝑑𝑢 = 1/8
+0.5 +0.5 −∞

-0.5 +0.5 -0.5 +0.5

convolution: properties

Cummutative operator

Associative operator

Distributive operator

- Convolution with dirac function:

convolution: link with Fourier transform

A filter can be defined in the frequency domain from the Fourier transform:

Let us define a system of impulse response h(t) and an input signal

The output is given by

+∞
(𝒚 𝒕 =න 𝒉 𝝉 𝒙 𝒕 − 𝝉 𝒅𝝉 )
−∞

Complex gain or Transfert function of the system

convolution: link with the Fourier transform

Consider a signal x(t) and its inverse Fourier transform:

Each component is corresponding to the output and by superposition

Then we deduce the Fourier transform

Convolution Product in the

⇋ In the time ⇋ frequency
domain domain
convolution: link with the Fourier transform

We have
Plancherel’s theorem

Proof
convolution: Plancherel’s theorem

We want to compute:

Take the Fourier transform

By computing the inverse Fourier transform, we obtain

convolution: digital version

Let us define : x=(x(1),x(2),…,x(k),…,x(N) ) a sampled signal and a filter h composed of M coefficients

: h=(h(1),h(2),…,h(j),…,h(M) ) with M<<N. Then the digital convolution (x*h)(n) is given by:

𝑀
𝐸[ 2 ]
1
𝑥∗ℎ 𝑛 = ෍ 𝑥 𝑛−𝑘 ℎ 𝑘 , 𝑛 = [0, … , 𝑁]
𝑀
𝑀
𝑘=−𝐸[ 2 ]

Remark: M will be odd (3,5,7,9,…) and bounds equals the lower integer part of M/2
𝒏=𝟐
Example:

𝒙 x(n-2) x(n-1) x(n) x(n+1) x(n+2) 𝑁=5

𝒉 𝒉(−1) ℎ(0) 𝒉(𝟏) 𝑀=3

𝑥 𝑛 + 1 ℎ −1 +𝑥 𝑛 ℎ 0 +𝑥 𝑛 − 1 ℎ +1

𝒌
convolution: example

𝑥(𝑛) ℎ(𝑛)

𝑛=0 𝑛=0

𝑥 ∗ ℎ 𝑛 : n=(1,2,3)

𝑥∗ℎ 𝑛 =1 𝑥∗ℎ 𝑛 =2 𝑥∗ℎ 𝑛 =3

convolution: boundaries

𝑥 ∗ ℎ 𝑛 : n=0? First option: zero padding

Management of the
extremities of the table? x 𝑥∗ℎ 𝑛
Second option: wrapping
𝒏=0

Third option: symmetry

fourth option: streching

convolution: Image

2D continuous version
Let us define x a continuous 2D signal and a 2D filter . Then the continuous 2D convolution 𝑥 ∗ 𝑦 𝜏, 𝜂
is defined by:

+∞ +∞
𝑥 ∗ ℎ 𝝉, 𝜼 = න න 𝑥 𝑢, 𝑣 ℎ 𝝉 − 𝑢, 𝜼 − 𝑣 𝑑𝑢 𝑑𝑣
−∞ −∞

2D digital version
Let us define x a discret time 2D signal of N*N samples and a 2D filter h of M*M coefficients, with M<<N.
2D digital convolution (x*h)(n,m) is given by:

𝑀 𝑀
𝐸[ 2 ] 𝐸[ 2 ]
1
𝑥 ∗ ℎ 𝒏, 𝒎 = 2 ෍ ෍ 𝑥 𝒏 − 𝑘, 𝒎 − 𝑙 ℎ 𝑘, 𝑙 , (𝑛, 𝑚) = [0, … , 𝑁]
𝑀
𝑀 𝑀
𝑘=−𝐸[ 2 ] 𝑙=−𝐸[ 2 ] Matlab : z=conv2(x,h)
convolution: example
1 1
1
M=3: 𝑥 ∗ ℎ 𝒏, 𝒎 = ෍ ෍ 𝑥 𝒏 − 𝑘, 𝒎 − 𝑙 ℎ 𝑘, 𝑙 ,
9
𝑘=−1 𝑙=−1

𝒎 Image area filter

x(n-1,m-1) x(n-1,m) x(n-1,m+1) h(-1,-1) h(-1,0) h(-1,1)

𝒏
x(n, m-1) x(n,m) x(n,m+1) h(0, -1) h(0,0) h(0,1)

x(n+1, m-1) x(n+1,m) x(n+1,m+1) h(1, -1) h(1,0) h(1,1)

𝑥 ∗ ℎ 𝒏, 𝒎 = x(n+1,m+1).h(−1,−1)+x(n+1,m).h(−1,0)+ ⋯ .

𝟏/𝟗
convolution: example

(M=3)

original M=3 M=7

original