Chapter 04c Frequency Filtering (Circulant Matrices) 1spp

Digital Image Processing
Filtering in the Frequency Domain

(Circulant Matrices and Convolution)
Christophoros Nikou
cnikou@cs.uoi.gr
University of Ioannina - Department of Computer Science
Toeplitz matrices
Elements with constant value along the main

diagonal and sub-diagonals.
For a NxN matrix, its elements are determined by
a (2N-1)-length sequence {tn | ( N 1) n N 1}
T(m, n) = tm n
t0
t
1
T = t2
N 1
t1 t2 t( N 1)
t0 t1
t2
t1
t2 t1 t0 NN
C. Nikou Digital Image Processing (E12)
Toeplitz matrices (cont.)
Each row (column) is generated by a shift of the

previous row (column).
The last element disappears.
A new element appears.
T(m, n) = tm n
t0
t
1
T = t2
N 1
t1 t2 t( N 1)
t0 t1
t2
t1
t2 t1 t0 NN
Circulant matrices
Elements with constant value along the main
diagonal and sub-diagonals.
For a NxN matrix, its elements are determined by
a N-length sequence {cn | 0 n N 1}
c1 c2
c0
c
C(m, n) = c( m n ) mod N
c
c
0
1
N 1
C = cN 2
c1
c2 cN 1
cN 1
c2
c1
c0 NN
Circulant matrices (cont.)

Special case of a Toeplitz matrix.
Each row (column) is generated by a circular shift
(modulo N) of the previous row (column).
c1 c2
c0
c
C(m, n) = c( m n ) mod N
c
c
0
1
N 1
C = cN 2
c1
c2 cN 1
cN 1
c2
c1
c0 NN
Convolution by matrix-vector
operations
1-D linear convolution between two discrete
signals may be expressed as the product of a
Toeplitz matrix constructed by the elements of
one of the signals and a vector constructed by the
elements of the other signal.
1-D circular convolution between two discrete
signals may be expressed as the product of a
circulant matrix constructed by the elements of
one of the signals and a vector constructed by the
elements of the other signal.
Extension to 2D signals.
1D linear convolution using Toeplitz

matrices
f [n] = {1, 2, 2}, h[n] = {1, 1}, N1 = 3, N 2 = 2
The linear convolution g[n]=f [n]*h [n] will be of

length N=N1+N2-1=3+2-1=4.
We create a Toeplitz matrix H from the
elements of h [n] (zero-padded if needed) with
N=4 lines (the length of the result).
N1=3 columns (the length of f [n]).
The two signals may be interchanged.

matrices (cont.)
f [n] = {1, 2, 2}, h[n] = {1, 1}, N1 = 3, N 2 = 2
1 0 0
1 1 0
H=
0 1 1
0 0 1 43
Length of f [n] = 3
Zero-padded h[n] in the first column
Length of the result =4

Notice that H is not
circulant (e.g. a -1
appears in the second line
which is not present in the
first line.

matrices (cont.)
f [n] = {1, 2, 2}, h[n] = {1, 1}, N1 = 3, N 2 = 2
1 0 0
1
1 1 0 1 1

g = Hf =
=
2
0 1 1 0
2
0 0 1
2
g[n] = {1, 1, 0, 2}
10
1D circular convolution using

circulant matrices
f [n] = {1, 2, 2}, h[n] = {1, 1}, N1 = 3, N 2 = 2
The circular convolution g[n]=f [n]h [n] will

be of length N=max{N1, N2}=3.
We create a circulant matrix H from the
elements of h [n] (zero-padded if needed)
of size NxN.
The two signals may be interchanged.
11

circulant matrices (cont.)
f [n] = {1, 2, 2}, h[n] = {1, 1}, N1 = 3, N 2 = 2
1 0 1
H = 1 1 0
0 1 1 33
Zero-padded h[n] in the
first column
12

f [n] = {1, 2, 2}, h[n] = {1, 1}, N1 = 3, N 2 = 2
1 0 1 1 1
g = Hf = 1 1 0 2 = 1
0 1 1 2 0
g[n] == { 1, 1, 0}
Block matrices
13
A11
A
A = 21
AM 1
A12
A22
AM 2
A1N
A2 N
AMN
Aij are matrices.

If the structure of A, with respect to its sub-matrices, is
Toeplitz (circulant) then matrix A is called block-Toeplitz
(block-circulant).
If each individual Aij is also a Toeplitz (circulant) matrix then
A is called doubly block-Toeplitz (doubly block-circulant).
14
2D linear convolution using doubly

block Toeplitz matrices
m
m
1 4 1
2 5 3
f [m,n]
M1=2, N1=3
1 1
1 -1
h [m,n]
M2=2, N2=2
The result will be of size (M1+M2-1) x (N1+N2-1) = 3 x 4

15

block Toeplitz matrices (cont.)
m
m
1 4 1
2 5 3
f [m,n]
1 1
1 -1
h [m,n]
At first, h[m,n] is zero-padded

m
to 3 x 4 (the size of the
0 0 0 0
result).
1 1 0 0
Then, for each row of h[m,n],
1 -1 0 0
a Toeplitz matrix with 3
columns (the number of
h[m,n]
columns of f [m,n]) is
constructed.

16
m
0 0 0 0
1 1 0 0
1 -1 0 0
h[m,n]
1 0 0
1 1 0
H1 =
0 1 1
0 0 1
For each row of h[m,n], a

Toeplitz matrix with 3 columns
(the number of columns of f
[m,n]) is constructed.
1
1
H2 =
0
0
1
1
0
0
0
1
0
0
H3 =
0
0 0
0 0
0 0
0 0
17

m
m
1 4 1
2 5 3
1 1
1 -1
h [m,n]
f [m,n]
Using matrices H1, H2
and H3 as elements, a
doubly block Toeplitz
matrix H is then
constructed with 2
rows of f [m,n]).
H1
H = H 2
H 3
H3
H1
H 2 126
18

m
m
1 4 1
2 5 3
f [m,n]
We now construct
a vector from the
elements of f [m,n].
1 1
1 -1
h [m,n]
2
5

3 (2 5 3)T
f = =
T
1 (1 4 1)
4

1
19

m
m
1 4 1
2 5 3
1 1
1 -1
h [m,n]
f [m,n]
H1
g = Hf = H 2
H 3
2
5
H3
3
H1
1
H 2
4

1
20

1 0 0 0 0 0
2
1 1 0 0 0 0
3

0 1 1 0 0 0
2

0
0
1
0
0
0
2
3
1 0 0 1 0 0 5 3
(2 3 2 3)T
1 1 0 1 1 0 3 10
T
g = Hf =
= = (3 10 5 2)
0 1 1 0 1 1 1
5
(1 5 5 1)T
0 0 1 0 0 1 4 2
0 0 0 1 0 0 1 1

0 0 0 1 1 0
5

0
0
0
0
1
1
5
0 0 0 0 0 1
1

21
m
1 4 1
2 5 3
1 1
1 -1
h [m,n]
f [m,n]
m
1 5 5 1
3 10 5 2
2 3 -2 3
(2 3 2 3)T
T
g = (3 10 5 2)
(1 5 5 1)T
g [m,n]
22

Another example
m
m
3 4
1 2
f [m,n]
M1=2, N1=2
1 -1
h [m,n]
M2=1, N2=2
The result will be of size (M1+M2-1) x (N1+N2-1) = 2 x 3

23

m
m
3 4
1 2
f [m,n]
1 -1
h [m,n]
At first, h[m,n] is zero-padded

m
to 2 x 3 (the size of the
result).
0 0 0
Then, for each line of h[m,n],
1 -1 0
a Toeplitz matrix with 2
h[m,n]
columns of f [m,n]) is
constructed.

24
m
0 0 0
1 -1 0
h[m,n]
1 0
H1 = 1 1
0 1
For each row of h[m,n], a Toeplitz

matrix with 2 columns (the
number of columns of f [m,n]) is
constructed.
n
0 0
H 2 = 0 0
0 0
25

m
m
3 4
1 2
1 -1
h [m,n]
f [m,n]
Using matrices H1 and
H2 as elements, a
doubly block Toeplitz
matrix H is then
constructed with 2
rows of f [m,n]).
H1
H=
H 2
H2
H1 64
26

m
m
3 4
1 2
f [m,n]
We now construct a
vector from the elements
of f [m,n].
1 -1
h [m,n]
1
2 (1 2)T
f = =
T
3 (3 4)

4
27

m
m
3 4
1 2
f [m,n]
H1
g = Hf =
H 2
1 -1
h [m,n]
1
H 2 2

H1 3

4
28

1 0 0 0
1
1 1 0 0 1
1

0 1 0 0 2 2 (1 1 2)T
g = Hf =
= =
T
0 0 1 0 3 3 (3 1 4)
0 0 1 1 4 1

0 0 0 1
4

29
3 4
1 2
1 -1
h [m,n]
f [m,n]
m
3 1 4
1 1 -2
(1 1 2)T
g=
T
(3 1 4)
g [m,n]
30
2D circular convolution using doubly

block circulant matrices
The circular convolution g[m,n]=f [m,n]h [m,n]

with 0 m M 1, 0 n N 1,
may be expressed in matrix-vector form as:
g = Hf
where H is a doubly block circulant matrix
generated by h [m,n] and f is a vectorized form
of f [m,n].
31

block circulant matrices (cont.)
H0
H
1
H = H2
HM 1
HM 1
H0
H1
HM 2
HM 2 H1
HM 1 H2
H0
H3
HM 3 H0
Each Hj, for j=1,..M, is a circulant matrix with N

columns (the number of columns of f [m,n])
generated from the elements of the j-th row of
h [m,n].
32

h[ j, N 1] h[ j,1]
h[ j,0]
h[ j,1]
h
[
j
,0]
h
[
j
,
2]
Hj =
h[ j, N 1] h[ j, N 2] h[ j,0] NN
Each Hj, for j=1,..M, is a NxN circulant matrix

generated from the elements of the j-th row of
h [m,n].
33

m
m
0 1 0
1 3 -1
1 2 1
f [m,n]
1 0 1
H 0 = 1 1 0
0 1 1
1 0 0
H1 = 0 1 0
0 0 1
0 0 0
1 0 0
1 -1 0
h [m,n]
0 0 0
H 2 = 0 0 0
0 0 0
34

m
m
0 1 0
1 3 -1
1 2 1
f [m,n]
H0
g = Hf = H1
H 2
H2
H0
H1
0 0 0
1 0 0
1 -1 0
h [m,n]
H1 (1 2 1)
T
H 2 (1 3 1)
H 0 ( 0 1 0 )T
35

1 0 1 0 0 0 1 0 0 1 0
1 1 1 0 0 0 0 1 0 2 2

0 1 0 0 0 0 0 0 1 1 1

1
0
0
1
0
1
0
0
0
1 3
g = Hf = 0 1 0 1 1 1 0 0 0 3 = 4

0 0 1 0 1 0 0 0 0 1 3
0 0 0 0 0 0 1 0 1 0 1

0 0 0 0 0 0 1 1 1 1 4
0 0 0 0 0 0 0 1 0 0 2
36

m
m
0 1 0
1 3 -1
1 2 1
f [m,n]
1 4 -2
3 4 -3
0 2 -1
0 0 0
1 0 0
1 -1 0
h [m,n]
( 0 2 1)T
T
g = ( 3 4 3)
T
(1 4 2 )
g [m,n]
37
Diagonalization of circulant matrices

Theorem: The columns of the inverse DFT matrix
are eigenvectors of any circulant matrix. The
corresponding eigenvalues are the DFT values of
the signal generating the circulant matrix.
Proof: Let
wN = e
2
j
w =e
nk
N
2 n
j
k
be the DFT basis elements of length N with:
0 k N 1, 0 n N 1,
38

(cont.)
We know that the DFT F [k] of a 1D signal f [n]
may be expressed in matrix-vector form:
F = Af
where
f = [ f [0], f [1],..., f [ N 1]] , F = [ F [0], F [1],..., F [ N 1]]

T
( w0 ) 0
N
1 0
(w )
A= N
( w N 1 )0
N
(w )
(w )
0 1
N
1 1
N
(w )
(w )
0
N
1
N
(w ) (w )
N 1 1
N
N 1 2
N
(w )
(w )
1 N 1
1
N
( wNN 1 )
0
N
N 1

(cont.)
39
The inverse DFT is then expressed by:

-1
f =A F
where
1
* T
A = (A )
N
1
0
( wN )
1 0
1 ( wN )
=
N
N 1 0
( wN )
( wN0 )
( wN0 )
(w )
(w )
1 1
N
1
N
(w ) (w )
N 1 1
N
N 1 2
N
( wN0 )
1 N 1
( wN )
N
1
( wNN 1 )
N 1
* T
The theorem implies that any circulant matrix has

eigenvectors the columns of A-1.
40

(cont.)
Let H be a NxN circulant matrix generated by the 1D

N-length signal h[n], that is:
H (m, n) = h [ (m n) mod N ] h[m n]N

Let also k be the k-th column of the inverse DFT
matrix A-1.
We will prove that k, for any k, is an eigenvector of H.
The m-th element of the vector Hk, denoted by [ H k ]m

is the result of the circular convolution of the signal
h[n] with k.
41

(cont.)
[ H k ]m
l =mn
1
=
N
N 1
1
= h [ m n ]N k [ n] =
N
n =0
m ( N 1)
l =m
k ( m l )
N
h [l ]N w
N 1
h [ m n]
n=0
1 k m
= wN
N
k n
N
m ( N 1)
l =m
h [l ]N wNk l
m
1 k m 1
kl
kl
= wN h [l ]N wN + h [l ]N wN
N
l =0
l = m ( N 1)
We will break it into two parts

42

(cont.)
N 1
N 1
1 k m 1
kl
kl
kl
= wN h [l ]N wN + h [l ]N wN h [l ]N wN
N
l =0
l = m +1
l = m ( N 1)
Periodic with respect to N.

N 1
N 1
1 k m N 1
kl
kl
kl
N
l =0
l = m +1
l = N + m ( N 1)
N 1
N 1
1 k m N 1
kl
kl
kl
N
l =0
l = m +1
l = m +1
43

(cont.)
[ H k ]m
1 k m N 1
kl
= wN h [l ]N wN
N
l =0
= H [k ][ k ]m
DFT of h[n] at k.
This holds for any value of m. Therefore:
H k = H [k ] k
which means that k, for any k, is an eigenvector of
H with corresponding eigenvalue the k-th element
of H[k], the DFT of the signal generating H .
44

(cont.)
The above expression may be written in terms of
the inverse DFT matrix:
1
HA = A
=diag { H [0], H [1],..., H [ N 1]}
or equivalently: = HA 1
Based on this diagonalization, we can prove the
property between circular convolution and DFT.
45

(cont.)
g = Hf g = HA -1 Af
Ag = AHA -1 Af G = F
0
0
G[0] H [0]
F [0]
G[0] 0
H
[1]
0
F
[0]
=
0 H [ N 1] F [ N 1]
G[ N 1] 0
DFT of g[n]
DFT of h [n]
DFT of f [n]
Diagonalization of doubly block

circulant matrices
46
These properties may be generalized in 2D.

We need to define the Kronecker product:
A
M N
, B
K L
a11B a12 B a1N B

a B a B a B
22
2N
A B = 21
aM 1B aM 2 B aMN B MK NL
47

The 2D signal f [m,n], 0 m M 1, 0 n N 1,

may be vectorized in lexicographic ordering
(stacking one column after the other) to a
vector:
MN1
f
The DFT of f [m,n], may be obtained in
matrix-vector form:
F = (A A)f
48

Theorem: The columns of the inverse 2D DFT
matrix
1
(A A)
are eigenvectors of any doubly block circulant

matrix. The corresponding eigenvalues are the 2D
DFT values of the 2D signal generating the doubly
block circulant matrix:
= (A A) H (A A)
Diagonal, containing the 2D DFT
of h[m,n] generating H
Doubly block circulant

Chapter 04c Frequency Filtering (Circulant Matrices) 1spp

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 04c Frequency Filtering (Circulant Matrices) 1spp

Uploaded by

Copyright:

Available Formats

Digital Image Processing

Filtering in the Frequency Domain

University of Ioannina - Department of Computer Science

Elements with constant value along the main

C. Nikou Digital Image Processing (E12)

Toeplitz matrices (cont.)

Each row (column) is generated by a shift of the

C. Nikou Digital Image Processing (E12)

Circulant matrices (cont.)

1D linear convolution using Toeplitz

The linear convolution g[n]=f [n]*h [n] will be of

1D linear convolution using Toeplitz

Length of the result =4

C. Nikou Digital Image Processing (E12)

1D linear convolution using Toeplitz

1D circular convolution using

The circular convolution g[n]=f [n]h [n] will

C. Nikou Digital Image Processing (E12)

1D circular convolution using

1D circular convolution using

C. Nikou Digital Image Processing (E12)

Aij are matrices.

2D linear convolution using doubly

The result will be of size (M1+M2-1) x (N1+N2-1) = 3 x 4

2D linear convolution using doubly

At first, h[m,n] is zero-padded

2D linear convolution using doubly

For each row of h[m,n], a

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

2D linear convolution using doubly

2D linear convolution using doubly

The result will be of size (M1+M2-1) x (N1+N2-1) = 2 x 3

2D linear convolution using doubly

At first, h[m,n] is zero-padded

2D linear convolution using doubly

For each row of h[m,n], a Toeplitz

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly

2D circular convolution using doubly

The circular convolution g[m,n]=f [m,n]h [m,n]

2D circular convolution using doubly

Each Hj, for j=1,..M, is a circulant matrix with N

2D circular convolution using doubly

Each Hj, for j=1,..M, is a NxN circulant matrix

2D circular convolution using doubly

C. Nikou Digital Image Processing (E12)

2D circular convolution using doubly

C. Nikou Digital Image Processing (E12)

2D circular convolution using doubly