Chapter 04c Frequency Filtering (Circulant Matrices) 2spp

3/14/2011
Digital Image Processing
Filtering in the Frequency Domain

(Circulant Matrices and Convolution)
Christophoros Nikou Niko 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 t
N 1
t1 t2 t( N 1) t0 t1 t2 t1 t2 t1 t0 N N
C. Nikou Digital Image Processing (E12)
3/14/2011
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 t
N 1
t1 t2 t( N 1) t0 t1 t2 t1 t2 t1 t0 N N
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 N 1 c0 c1 C = cN 2 c2 cN 1 c1
cN 1 c2 c1 c0 N N
3/14/2011
Circulant matrices (cont.)

Special case of a Toeplitz matrix. (column) ) is g generated by y a circular shift Each row ( (modulo N) of the previous row (column).
C(m, n) = c( m n ) mod N
c1 c2 c0 c N 1 c0 c1 C = cN 2 c2 cN 1 c1
cN 1 c2 c1 c0 N N
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.
3/14/2011
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.
1D linear convolution using Toeplitz 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.
3/14/2011
1D linear convolution using Toeplitz matrices (cont.)

f [n] = {1, 2, 2}, h[n] = {1, 1}, N1 = 3, N 2 = 2
1 0 0 1 1 1 1 0 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 f size i NxN.
The two signals may be interchanged.
3/14/2011
11
1D circular convolution using circulant matrices (cont.)

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

f [n] = {1, 2, 2}, h[n] = {1, 1}, N1 = 3, N 2 = 2
1 0 1 1 1 2 = 1 1 1 0 g = Hf = 0 1 1 2 0
g[n] == { 1, 1, 0}
3/14/2011
13
Block matrices
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 1 4 1 2 5 3 f [m,n] M1=2, N1=3 m 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

3/14/2011
15
2D linear convolution using doubly block Toeplitz matrices (cont.)

m 1 4 1 2 5 3 f [m,n] m 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. C. Nikou Digital Image Processing (E12)
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
0 0 1 0 1 1 0 1
0 0 H3 = 0 0
0 0 0 0 0 0 0 0
3/14/2011
17

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

m 1 4 1 2 5 3 f [m,n] m 1 1 1 -1 h [m,n]
2 5 3 (2 5 3)T f = = T (1 4 1) 1 4 1
We now construct a vector from the elements of f [m,n].
3/14/2011
19

m 1 4 1 2 5 3 f [m,n]
H1 H g = Hf = 2 H3 2 5 H3 3 H1 1 H2 4 1
m 1 1 1 -1 h [m,n]
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 T (2 3 2 3) 1 1 0 1 1 0 3 10 T g = Hf = 0 1 1 0 1 1 1 = 5 = (3 10 5 2) (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
10
3/14/2011
21

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

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

11
3/14/2011
23

m 3 4 1 2 f [m,n] m 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 columns (the number of h[m,n] columns of f [m,n]) is constructed. C. Nikou Digital Image Processing (E12)
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 H2 = 0 0 0 0
12
3/14/2011
25

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

m 3 4 1 2 f [m,n] m 1 -1 h [m,n]
1 2 (1 2)T f = = T 3 (3 4) 4
We now construct a vector from the elements of f [m,n].
13
3/14/2011
27

m 3 4 1 2 f [m,n] m 1 -1 h [m,n]
H g = Hf = 1 H 2
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 (3 1 4) 0 0 1 0 3 3 0 0 1 1 4 1 0 0 0 1 4
14
3/14/2011
29

m 3 4 1 2 f [m,n] m m
*
n
1 -1 h [m,n]
3 1 4 1 1 -2 g [m,n]
(1 1 2)T g= T (3 1 4)
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, 1 0 n N 1, 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].
15
3/14/2011
31
2D circular convolution using doubly block circulant matrices (cont.)

H0 H 1 H = H2 HM 1 HM 1 H0 H1 H M 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

0] h [ j, N 1] h [ j,1] 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].
16
3/14/2011
33

m m 0 1 0 1 3 -1 1 2 1 f [m,n]

n
0 0 0 1 0 0 1 -1 0 h [m,n]
1 0 1 H0 = 1 1 0 0 1 1
1 0 0 H1 = 0 1 0 0 0 1
0 0 0 H2 = 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

n
0 0 0 1 0 0 1 -1 0 h [m,n]
T (1 2 1) H1 T (1 3 1) H2 T H0 (0 1 0)
17
3/14/2011
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]

n
0 0 0 1 0 0 1 -1 0 h [m,n]
1 4 -2 3 4 -3 0 2 -1 g [m,n]
( 0 2 1)T T g = ( 3 4 3) T 1 4 2 ) (
18
3/14/2011
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
2 2 n k
wN = e
w =e
nk N
be the DFT basis elements of length N with:
0 k N 1, 0 n N 1,
38
Diagonalization of circulant matrices (cont.)

We know that the DFT F [k] of a 1D signal f [n] may be expressed in matrix-vector matrix vector form:
F = Af
where
f = [ f [0], f [1],..., f [ N 1]] , F = [ F [0], F [1],..., F [ N 1]]
T 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
1 N 1 N N 1 N 1 ( wN )
0 N N 1
(w ) (w )
19
3/14/2011
39

The inverse DFT is then expressed by:
f = A -1F
where
T 1 A = ( A* ) N 1
0 0 ( wN ) 0 1 ( w1 N ) = N N 1 0 wN ( )
(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 ) 1 N 1 ( wN ) N 1 N 1 ( wN )
0 N N 1
* T
The theorem implies that any circulant matrix has eigenvectors the columns of A-1.
40

H (m, n) = h [ (m n) mod N ] h[m n]N
Let H be a NxN circulant matrix generated by the 1D N-length signal h[n], that is:
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. is the result of the circular convolution of the signal h[n] with k.
The m-th element of the vector Hk, denoted by [ H k ]m
20
3/14/2011
41

1 = h [ m n]N k [ n ] = N n =0
N 1
[ H k ]m
l =mn
h [ m n]
n =0
N 1
k n wN
1 = N
m ( N 1)
l =m
h [l ]N w
k ( m l ) N
1 k m m ( N 1) kl = wN h [l ]N wN N l =m
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

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 = wN h [l ]N wN + h [l ]N wN h [l ]N wN N l =0 l = m +1 l = N + m ( N 1)
N 1 N 1 1 k m N 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 +1
21
3/14/2011
43

1 k m N 1 kl wN h [l ]N wN N l =0
DFT of h[n] at k.
[ H k ]m =
= H [k ][ k ]m
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

The above expression may be written in terms of the inverse DFT matrix:
HA 1 = A 1
=diag { H [0], H [1],..., H [ N 1]}
or equivalently: q y = HA 1 Based on this diagonalization, we can prove the property between circular convolution and DFT.
22
3/14/2011
45

Ag = AHA -1 Af G = F
g = Hf g = HA -1 Af
0 0 G[0] H [0] F [0] G[0] 0 F [0] H [1] 0 = 0 H [ N 1] F [ N 1] G[ N 1] 0

DFT of g[n] DFT of h [n] DFT of f [n]
46
Diagonalization of doubly block circulant matrices
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 B B B a a a M2 MN MK NL M1

23
3/14/2011
47
Diagonalization of doubly block circulant matrices (cont.)

may be vectorized in lexicographic ordering (stacking one column after the other) to a vector: f MN 1
The 2D signal f [m,n], 0 m M 1, 0 n N 1,
The DFT of f [m,n], may be obtained in matrix-vector form:

F = ( A A)f
48
Diagonalization of doubly block circulant matrices (cont.)

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
24

Chapter 04c Frequency Filtering (Circulant Matrices) 2spp

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) 2spp

Uploaded by

Copyright:

Available Formats

3/14/2011

Digital Image Processing

Filtering in the Frequency Domain

University of Ioannina - Department of Computer Science

C. Nikou Digital Image Processing (E12)

Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

Circulant matrices (cont.)

C. Nikou Digital Image Processing (E12)

Convolution by matrix-vector operations

1D linear convolution using Toeplitz matrices

1D linear convolution using Toeplitz matrices (cont.)

Length of the result =4

C. Nikou Digital Image Processing (E12)

1D linear convolution using Toeplitz matrices (cont.)

1D circular convolution using circulant matrices

C. Nikou Digital Image Processing (E12)

1D circular convolution using circulant matrices (cont.)

1D circular convolution using circulant matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices

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

2D linear convolution using doubly block Toeplitz matrices (cont.)

2D linear convolution using doubly block Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

We now construct a vector from the elements of f [m,n].

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

2D linear convolution using doubly block Toeplitz matrices (cont.)

(2 3 2 3)T g = (3 10 5 2)T (1 5 5 1)T

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

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

2D linear convolution using doubly block Toeplitz matrices (cont.)

2D linear convolution using doubly block Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

We now construct a vector from the elements of f [m,n].

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D linear convolution using doubly block Toeplitz matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D circular convolution using doubly block circulant matrices

2D circular convolution using doubly block circulant matrices (cont.)

2D circular convolution using doubly block circulant matrices (cont.)

2D circular convolution using doubly block circulant matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D circular convolution using doubly block circulant matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D circular convolution using doubly block circulant matrices (cont.)

C. Nikou Digital Image Processing (E12)

2D circular convolution using doubly block circulant matrices (cont.)

C. Nikou Digital Image Processing (E12)

Diagonalization of circulant matrices