Robust Control (Example)
Robust Control (Example)
Aug., 2005
ABSTRACT
There are many common methods in system analysis, design and development in two different
fields of control theory and texture image processing. The progress of these two area also shows
that the techniques developed to solve the problems of one area often find applications to the
other one.
This proposal is an attempt to setup and solve the texture image processing problem from
a viewpoint of control, especially the study of robust identification and robust control theory.
The work will focus on texture modelling, synthesis, recognition and classification. A novel
image modelling and model reduction approach is introduced. We will show how the recently
developed robust identification techniques can be applied to find models for textures which are
capable to do image compression and reconstruction. The work on extending the current 1-D
approach to 2-D to remove the column and row oriented limitation are mainly discussed in this
chapter.
Furthermore, based on the models discussed in the prior chapter, we apply semi-blind model
(in)validation method to address the problem of texture recognition and classification. The work
was motivated by the recent progress in the model (in)validation of robust control. A LMI based
convex relaxation was introduced to the original NP hard minimization problem and an efficient
solution was brought out. Under the proposed framework, texture recognition and classification
problem was recast into a robust model (in)validation form and was solved by introducing the
similar convex relaxation.
Moreover, a Hankel operator based approach to the problems of texture modelling and
inpainting is introduced. Textured images are modelled as the output of an unknown, periodic,
linear shift invariant operator in response to a suitable input. This approach can be applied to
missing portions reconstruction, finding textons and synthesizing textures from the same family.
It has great potential in the 2-D extension using the 2-D models and Hankel matrix indicated
in this proposal.
Keywords:
Contents
1 Introduction
1.1
1.2
Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1
2.2
2.2.1
2.2.2
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
12
2.3.1
12
2.3.2
13
2.3.3
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
2.3
20
3.1
20
3.2
21
3.3
23
3.3.1
23
3.3.2
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
3.4
. . . . . . . . .
29
3.4.1
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
3.4.2
31
33
4.1
Texture modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.2
Finding textons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.3
Texture Inpaiting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
37
Introduction
Control theory is a comprehensive research area with great power for the applications to wide
fields. Robust control and robust identification as the theories which highlight and deal with the
model uncertainty problem got considerable attentions and have been quickly developed during
recent decades. Signal processing is also an exciting area of research. Much recent progress has
been driven by applications involving speech and audio processing, image processing and others.
The research in this proposal arises in the context of many practical problems in different fields,
where the use of control-theoretical tools can bring in great benefit. It will give sufficient solution
to many problems, like image restoration, texture recognition and texture inpainting, from a
novel viewpoint of control theory. The following chapters will give more detailed explanation to
the theories and applications.
1.1
niques can be brought to bear on the problem of texture image processing involving image
modelling, texture synthesis, texture recognition and classification.
The proposal is organized as follows.
Part 1 gives the general introduction and the organization of the proposal. The notations
used throughout the proposal are also given in this part.
Part 2 gives a summary of the undergoing research progress in texture modelling. A finite
horizon model reduction method is introduce. There are two methods indicated in this chapter
to extend the finite horizon modelling and reduction method from 1-D to 2-D: 2-D filter SVD
realization algorithm and a new realization method for Roessers image processing model.
In part 3, the semi-blind model (in)validation theory is introduced. The problem setup,
transform and an efficient relaxation for the applications of model (in)validation method to
texture synthesis was discussed. Several examples with the discussion will give some light to the
efficiency and performance of this algorithm.
In part 4, a Hankel operator approach to texture modelling and inpaiting is introduced. The
image restoration and in painting problem is converted to a rank minimization problem.
Finally, the conclusion and the future work leading to my Ph.D degree are discussed in part
5.
1.2
Notations
Below we summarize the notation used in this proposal:
Z, R, C: set of integer, real and complex numbers respectively.
x, x : complex-valued column vector and its conjugate transpose row vector.
||x||: Euclidean norm of vector x Cm .
1
. P
p p
||x||p : p-norm of a vector, ||x||p = ( m
k=1 |xk | ) .
P
iZ
Xi z 1 .
standing problems in computer vision [15] with applications to widely dissimilar areas.
The first general texture model was proposed by Julesz [6] who suggested that texture
perception could be explained by extracting k th order statistics. However, the amount of data
contained in high order statistics, which is normal for human visual systems, is huge and difficult
to handle by computer algorithms. Most recent statistical approaches are based on two well
established areas. The first one is the theory of Markov random fields, which is used to model
statistical local interactions within small neighborhoods [711]. The second one is the use of
multiple linear kernels at different scales and orientations followed by a non-linear procedure,
motivated by the multi-channel filtering mechanisms generally accepted by neurophysiology
[1217].
Texture synthesis algorithms can be classified as procedural-based algorithms. Most procedural approaches are based on the statistical approaches described above. However, the cost of
tuning parameters of the models can be a limiting factor to synthesize large sets of textures. On
the other hand, image-based algorithms work by directly copying image pixels or patches from
a sample texture image and stitching them together in the synthesized image.
Efros and Leung [18] grow texture by coping pixel from a sample image using a statistical
non-parametric model based on the distribution of the brightness values in the neighborhood of
the pixel. Many optimization work to the basic algorithm has been done, such as Wei and Leung
[19], Harrison [20] and Ashikhmin [21]. Rane et al [22] address the problem of filling-in blocks of
missing data in wireless image transmission where the correlation between the lost data and its
neighbors is used to choose between an inpainting [23] and a texture synthesis [18] algorithm.
Pixel-based algorithms are greedy and can fail to capture the overall structure of the
texture. This problem can be alleviated by quilting patches instead. Xu et al. [24] use tiles
where the boundaries are obscured by pasting and blending random overlapping blocks. Efros
and Freeman in [25] stich together patches in a way to minimize the overlap error between
adjacent patches can have arbitrary shape and dimension. Liu et al [26],[27] generate nearregular textures by finding first its underlying lattice from a sample patch using concepts from
crystallographic symmetry groups.
The following portion will introduce a novel model reduction approach, its application to
2.1
of efficient algorithms are available [2832]. However the proposed approach to model reduction
of unstable systems cannot guarantee that certain structural properties, such as periodicity of
the impulse response, are preserved, a key requirement in the applications of special interests
in texture synthesis. Motivated by some well known results on realization theory [3334], a
different approach to model reduction is proposed by Sznaier [36]. The main idea is: Addressing
the problem by working directly with the constructed Hankel Matrix. Since the impulse response
is periodic, the Hankel matrix of the system under consideration is circulant and structural
properties can then be exploited to obtain balanced realizations in an efficient way.
The following is the brief introduction of the model reduction method and the application.
Consider a system Gn with the k th Markov parameter Mk = CAk1 B Rpm . The state
space realization of system Gn has the form: A Rnn , B Rnm , C Rpn and D
Rpm where An = I. The following algorithm is presented in [36] to get a rth order reduced
approximation Gr , such that ||Gn Gr || is minimized over the finite interval [0,n-1].
Algorithm 1:
1. Given the system with Mk = CAk1 B Rpm , form the np nm block matrix:
.
Hn =
M1 M2 Mn
M2 M3 M1
..
..
..
..
.
.
.
.
Mn M1 Mn1
"
#"
S 0
VT
Hn = U U
0 0
VT
S = diag(1 , , n ), i j , i j
3. Form the reduced order model:
Sr = diag(1 , , r ), r n
1
(1)
1
2
Cr = Ur Sr , Dr = D
where
P =
0
0
..
.
Ip
0
..
.
Ip
0 0
Ip 0
..
..
. .
0 0
(1)
Ur (Vr ) denotes the sub-matrix formed by the first r columns(rows) of Ur (Vr ) and where Ur
(1)
and Vr
Algorithm 1 can be applied to the non-trivial problems of texture synthesis and recognition,
algorithm 2. The main idea is to model images as the (periodic) impulse response of a nonnecessarily causal LTI system and use the proposed method to identify the corresponding model.
Partial images can be expended and additional realizations of the same texture can be obtained
by simply driving the corresponding model with a suitable input.
Algorithm 2:
1. Given an n m image, let RiT denote its ith row, and form the block matrix:
.
Hn =
R1 R2 Rn
R2 R3 R1
..
.. . .
..
.
.
.
.
Rn R1 Rn1
2.2
(2-D) recursive filtering with particular application to images. The operator theoretic approach
[36] considers images exhibiting a given static texture as realizations of one period of an infinite
2-D periodic signal, and then model this signal as the output of an unknown, periodic, linear
shift invariant operator with suitable input. However, this technique gets different model in
different row or column direction. Here we introduced a 2-D recursive filters SVD realization
to get the 2-D image model and remove this limitation. The 2-D periodic signal is considered
as the output of 2-D linear-phase IIR filters realized by a SVD-based structure. This structure
obtains a set of 1-D sub-filters which represent the model in two directions respectively.
2.2.1
Consider n m image as one period of an infinite 2-D signal with period (n,m). Therefore
9
the intensity value I(i, j) at a given pixel is satisfied with I(i, j) = I(i qn, j qm) for suitable
integer q. For simplicity, the length and width of the image are assumed to be the same, n = m.
It can be shown the realization approach for the case n 6= m can be easily extended.
Furthermore, this 2-D image signal is assumed as the output of 2-D IIR filters. Without
generality, we can assume the image is the responses to the specific case: input uk = (0),
i.e. the impulse response of the 2-D IIR filters. With these assumptions, the image modelling
problem becomes the realization of the impulse response of 2-D linear phase filters.
Motivated by the FIR filters realization algorithm [37], under the proposed image modelling
frame, for an image intensity value matrix I(i, j), 1 < i, j < n, the SVD decomposition to I is:
I = U SV
U2T
U1
"
0
0 0
#"
V1T
V2T
where U and V are unitary matrices; is a diagonal matrix with singular values in diagonal,
where 1 > 2 > > l > 0. V T is complicated conjugate of V .
Modify the SVD decomposition form to
I=
l
X
i ui viT
i=1
l
X
i ui , vi1 = i vi ,
i=1
Therefore 2-D FIR filer can be realized by l parallel 2-D sub-filters. Each 2-D sub-filter contains
two 1-D FIR filters in different row or column direction with the impulse response given by u1i
and vi1 .
This SVD technique can be used in model reduction by discarding the states small singular
values of I. Hence the reduced SVD structure can be expressed as I =
Plr
1 1T
i=1 ui vi .
We can
choose lr < l to get proper numbers of 2-D sub-filters in the filters realization.
For each 1-D FIR filter, with the assumption of periodic property, we can use the algorithm
2 to get an IIR realization. The IIR realization will be applied in texture synthesis and other
texture research.
Algorithm 3:
1. Given a n m image I(x, y), 1 < i, j < n, perform the SVD decomposition to I:
I = U SV
U1 U2T
"
0
0 0
#"
V1T
V2T
, = diag(1 , , l , )
l
X
i=1
i ui viT =
l
X
i=1
10
i ui , vi1 = i vi ,
3. Realize the 2-D FIR filter by l parallel 2-D sub-filters, with the form I =
Plr
1 1T
i=1 ui vi .
4. Apply Algorithm 2 and 1 on each 1-D sub-filter to obtain the reduced order realization.
2.2.2
Examples
Example 1:
For a 192 192 color image, we use 2-D filter SVD realization approach to get the texture
model with different rank of sub-filters, lr . The original image and outputs of different rank
2-D filters are showed in Figure 2.1. For 60 rank sub-filters, the output is slightly different from
the original image. Through the comparison we can conclude that the reduced 2-D SVD Filter
Realization model is efficient to describe the image.
Figure2.1 Original Image, Output of 40 2-D Filters, 60 2-D Filters, 80 2-D Filters
Example 2:
For different color or gray value images, use the 2-D IIR filters model to reconstruct partial
images. Figure 2.2 shows the efficiency of proposed algorithm in texture synthesis for specific
images.
From the above examples, we can get the conclusion that the modelling method can bead
to low order parsimonious models capable of generating images with the desired texture. It can
be used to both restore partial images and to classify an unknown sample.
11
2.3
limitation. Motivated by Roessers discrete state-space model for linear image processing [38],
an new image modelling approach is brought out here. We will first give the introduction to
Roessers model [38].
2.3.1
(1)
where
"
"
i
h
A1 A2
b1
A=
,B =
, C = c1 c2
A A4
b2
"
# "3
#
"
# "
#
x
(i, j)
R(i, j)
x
(i + 1, j)
R(i + 1, j)
.
.
0
x(i, j) =
=
, x (i, j) =
=
x
(i, j)
S(i, j)
x
(i, j + 1)
S(i, j + 1)
Roesser [36] and Takao [39] has shown that the solution of the model is given by
x(i, j) =
j
X
k=0
"
A(i,jk)
x
(0, k)
0
i
X
r=0
"
A(ir,j
0
x
(r, 0)
12
0<ki, 0<rj
M (k, r)u(i k, j r)
where
A(i,j)
(0,1)
M (k, r) = A"(k1,r)#b(1,0) + A(k,r1)
"
#b
b1
0
b(0,1) =
, b(0,1) =
0
b2
"
#
"
#
A1 A2
0
0
(1,0)
(0,1)
A
=
,A
=
0
0
A3 A4
(1,0)
(i1,j)
(0,1)
(i,j1)
=A
A
+A
A
, f or (i, j) > (0, 0); A(0,0) = I
With the assumption that A2 is a null matrix, which implies that the corresponding transfer
function has a separable denominator, the Markov parameters for the model can be generated
by
wi0 = c1 Ai1
1 b1
j1
woj = c2 A4 b2
i1
wij = c2 Aj1
4 A3 A1 b1
By taking the z-transform, we can compute the system function H(z1 , z2 ) = Y (z1 , z2 )/X(z1 , z2 )
given by
i
Y (z1 , z2 ) h
= c1 c2
H(z1 , z2 ) =
X(z1 , z2 )
"
z1 I 0
0 z2 I
"
A1 A2
A3 A4
#!1 "
b1
b2
+D
There is still problems to use a single Hankel matrix form and get A, B, C, D directly. In
[36], the algorithm to get the model uses the special forms of the Markov parameters. But for
2-D model, the Markov parameters have much more complex forms. It cant be expressed with
the form Mk = CAk1 B. And to generate future output needs not only current states, but also
the boundary conditions.
We get a system realization if and only if the system model can generate the same Markov
parameters sequence. Hence, applying the system modelling algorithm in [36] to Roessers
state-space model, we can solve the problem and get a 2-D model to generate the given Markov
parameters sequence.
2.3.2
Under the assumption that A2 is a null matrix, system (1) has the solutions
"
A(i,0)
{A(1,0) }i
Ai1 0
0 0
"
, A(0,j)
"
{A(0,1) }j
#
0
Aj1
4 A3
0
0
, i, j 1
j1
i 0
A
A
A
4
#3 1
"
#
"
0
b1
Ai1
1
, M (0, j) =
M (i, 0) =
0
Aj1
4 b2
#
"
0
, i, j 1.
M (i, j) =
i1
Aj1
A
3 A1 b1
4
A(i,j) =
13
0
Aj4
The results in paper [40], [41], [39] showed that the Hankel matrix generated by the realized
system can be written as the product of the observability and controllability matrices.
"
(n)
mm
H
0
(m)
Hnm = n,m n,m = H
nn
where
(m)
nn
H
(n)
(n) W
(n)
m
(m) W
(m) W
n(m)
W
W
W
1
2
1
2
W
(m) W
(m) W
(m)
(n) W
(n)
(n)
W
W
(n)
2
3
m+1
2
3
n+1
, Hmm =
..
..
..
..
..
..
..
..
.
.
.
.
.
.
.
.
(n)
(n)
(m)
(m)
(n)
(m)
Wn+1 Wn+n1
Wm Wm+1 Wm+m1
Wn
(m) = Cn Ai1 b1 = [wi0 , wi1 , , wim ]T
W
1
i
(n) = c2 Aj1 Bm = [w0j , w1j , , wnj ]
W
(m)
(n)
nn
mm
H
and H
are referred to as the horizontal and vertical Hankel matrices respectively. since
the given sequences assumed to have N as the least upper bound on its index, the n and m
can always be taken as N . Similar to the technique in [36], we form the horizontal and vertical
Hankel matrices as the following:
(N )
H
NN
(N ) W
(N ) W
(N
W
1
2
N
(N ) W
(N ) W
(N )
W
2
3
1
..
..
..
..
.
.
.
.
(N
(N
)
(N
)
)
W
W
W
1
N 1
N
(N
)
=
, H
NN
(N )
(N ) W
(N ) W
m
W
1
2
(N ) W
(N ) W
(N )
W
2
3
1
..
..
..
..
.
.
.
.
(N
(N
)
(N
)
)
W
W
W
1
N 1
N
Then using [36] algorithm, we get the system realization for horizontal and vertical Hankel
matrices respectively. The realization algorithm will give a realization solution set, including
CN , A1 , b1 , c2 , A4 and BN . Moreover, CN and BN can be expressed in the following forms.
CN =
c2 A3
c2 A4 A3
..
.
c2 A4N 1 A3
c2
c2 A4
..
.
.
A3 =
QN A3 ,
1
c2 AN
4
1
BN = [A3 b1 , A3 A1 b1 , , A3 AN
b1 ]
1
N 1
= A3 [b1 , A1 b1 , , A1 b1 ]
.
= A3 PN
or A3 = BN PN1
1 0
0 Q
A3 = (Q
N N ) QN CN
.
Wi0 =
w10
w20
..
.
wN 0
c2 b2
c2 A4 b2
..
.
c2 AN
4 b2
c2
c2 A4
..
.
N b2 = b2 = Q
1 Wi0
b2 = Q
c2 AN
4
0 Q
b2 = (Q
N N)
0 Wi0 .
Q
N
Similarly
W0j
.
=
w01
w02
..
.
i
i
h
h
= c1 b1 , c1 A1 b1 , , c1 AN b1 = c1 b1 , A1 b1 , , AN b1 = c1 PN
1
1
w0N
= c1 = W0j PN1
A=
2.3.3
A1 0
A3 A4
"
, B=
b1
b2
, C = [c1 , c2 ], D = w00
Examples
Example 1:
As a simple example, suppose the dimension of the interested image is 10 by 10, and the
15
W =
Wi0 =
52 91 128 80 72 60 82 62 64 48
158 165 146 95 72 49 53 65 59 60
91 128 80 72 60 82 62 64 48
W0j =
iT
, i = 1, , 9
, j = 1, , 9
Form the horizontal and vertical Hankel Matrices respectively, and apply [36] algorithms to
get the realization of CN , A1 , b1 , c2 , A4 and BN .
A1 =
1
0
0
0
0
0
0 0
0
0 0.76604 0.64279
0
0
0
0 0
0
0
0
0.76604
0
0
0
0 0
0
0
0
0
0.17365 0.98481
0
0 0
0
0
0
0
0
0.17365
0
0 0
0
0
0
0
0
0
0
0 0
0
0
0
0
0
0
0.86603 0 0
0
0
0
0
0
0
0
0 0 0.34202
0
0
0
0
0
0
0 0
0
h
b1 =
, A4 = A0
1
iT
c2 =
CN =
4.7056 0.75217 4.5448 2.0648 0.061258 0.11663
2.034
1.3527
2.2333
5.4888 0.68685 3.8443 1.7748 0.40345 1.8201
0.20412
1.803
2.4825
5.8448
0.3057
4.3093 1.9673
0.60427
1.1888
0.48354 0.024403 4.0105
7.7347
2.071
1.6779
2.6395
0.65053
3.2309
3.7108
0.37095
0.32521
5.7218 0.10584
3.4401
2.1262 0.016862 0.20599
1.2203
0.15129
1.0556
6.2719 0.51849 3.6301
0.766
3.6348
0.3508
3.8139
0.12966
0.19734
5.0551 2.0568
2.8378
1.591 0.50817
2.0402
0.53473 2.4311
1.818
5.1328 1.624
3.098
3.1573 3.1945 0.16196 1.2458
0.66318
2.1026
4.9451 1.9562
2.6829 2.5048
1.555
0.67568 1.1184
0.32847
2.3932
16
A3 =
0.023346 0.052747
0
0.00017604 0.025776 0.034817
0.16725
0
0.095289
0
0.030537 0.091546
0
0
0
0
0
0.055846
0
0
0
0
0
0
0.010131 0.11842
0.16197
0.21877
0
0.26139
0.074655
0
0.060109
0
0.037931
0
0
0
0
0
0.065239
0
0
0.029262
0
0.014859
0
0
0.33068 0.023437
0
0
0.021322 0.27673
0.060555
0
0.23893
0
00.012825
0
0.033871
0
0
0
0.084206
0
0
0
0
0
0.066208 0.054259
0
0
0.16721
b2 =
c1 =
A=
A1 0
A3 A4
"
R1818 , B =
b1
b2
Verify the realization result. The system can generate the identical Markov sequences to the
given image.
wi0 = c1 Ai1
1 b1
j1
woj = c2 A4 b2
i1
wij = c2 Aj1
4 A3 A1 b1
Here we get the full realization for the state-space model, and we can get reduced realization by
choose proper sub-matrix size r. If we choose r = 5, 7, 9, we get the following output:
Woriginal =
52 91 128 80 72 60 82 62 64 48
158 165 146 95 72 49 53 65 59 60
iT
Wr=8 =
Wr=5 =
52
155.76
135.83
93.806
51.415
78.415
69.806
47.829
54.756
116.39
88.874
141.25
97.702
81.86
52.036
42.145
71.856
54.255
57.134
130.51
129.73
170.21
120.29
84.561
86.426
55.176
41.674
65.222
85.88
137.07
80.869
47.62
76.46
88.058
164.74
186.91
98.649
104.01
136.47
67.03
63.733
57.71
84.505
90.391
101.27
141.04
106.87
52.863
71.808
73.031
61.874
67.009
51.763
66.769
147.24
179.36
112.66
56.336
51.814
61.847
50.263
85.399
67.832
63.905
122.42
139.03
83.934
59.455
66.445
73.7
52
158.16
139.12
84.927
61.296
72.533
70.685
48.541
56.35
112.39
93.759
130.32
100.52
70.966
70.303
74.118
57.218
44.105
70.087
116.54
115.39
159.62
137.93
94.165
58.268
40.902
38.609
54.567
93.625
139.68
78.207
54.196
59.115
101.35
140.29
139.76
114.63
99.681
95.291
78.187
71.743
64.984
57.844
90.759
142.69
151.99
106.98
63.533
61.373
73.46
50.655
68.386
54.83
78.912
136.09
158.87
115.7
60.284
51.062
70.148
56.596
87.131
64.821
67.621
107.44
130.45
100.45
57.83
57.127
84.329
Example 2:
For a 51 51 image, using full and reduced realization algorithms to extract the discrete
state-space model, and compare the result of the image synthesis with the models respectively.
Pictures in Figure 2.3 are respectively the fully realized image (equally to the sub-matrix
size r = 51), reduced image with r = 40, r = 30, r = 25, r = 15 and r = 10.
18
Figure2.3
Original image(r=51), reduced realization r=40, r=30, r=25, r=15 and r=10.
The approximation error analysis and other further theoretical analysis will be one of the
future work.
19
3.1
20
.
M22 (ej ) = 1 s(ej ) P22 (ej )Su (ej )ej()
The model (in)validation set-up form is as Figure 3.1 in [35].
In this frame work, the semi-blind model (in)validation problem can be precisely stated as
follows.
Problem 1: Given the output s(ej ) and the admissible sets of inputs U and noise N ,
determine whether there exists at least one pair z N , BH and a scalar function ()
so that system equations hold; or equivalently, whether:
= min, ||M ||2 1
The difficulty in solving problem 1 stems from the fact that the problem is not jointly convex
in , . To solve this problem, a tight convex relaxation of this problem is introduced and a
sufficient condition for solving problem 1 is proposed.
3.2
where
Maug
0
0
1
.
j
j
0
P11 (e )
P12 (e )Su (ej )
=
1
1
j
j
j
j
s(e ) P21 (e ) P22 (e )Su (e )
"
aug
.
=
ej() 0
0
21
"
.
Maug =
M11 M12
M21 M22
.
=
0
0
|
1
0
P11 (ej )
|
P12 (ej )Su (ej )
1
j
P21 (ej )
|
1 P22 (ej )Su (ej )
s(e )
2 BH
Note that the set aug is not convex due to the constrain |1 | = 1. To address the difficulty
and obtain a tractable optimization problem, we will relax the constrain to ||1 || 1. This leads
to the following model (in)validation problem with 2-block LTI structured uncertainty.
st = min
" st ||Maug st ||2 #
1 (ej )
0
.
st =
: ||i || 1
0
2 (ej )
Figure3.2 shows the transform of the semi-blind model (in)validation problem.
Figure3.2
The necessary and sufficient condition equivalent to st > 1 is is given in [35] as the following:
Consider a system M (z) RH and 2-block structured uncertainty st = {diag(1 , 2 ) :
|||| 1}. Then the following conditions ar equivalent:
(i) infst ||M ||22 > 1
(ii) There exists a Hermitian matrix X() 0 and a real transfer function y() 0, such
that in [0, 2) the following two inequalities hold:
"
X() 0
M (ej )
0
1
where X() = diag(x1 ()I1 , x2 ()I2 ).
R
02 y() d
2 > 1.
.
L(X, y) = M (ej )
22
"
X()
0
0
y()
0,
3.3
image processing communities, with application ranging from medical diagnosis to object recognition and image database retrieval. Discussed in the priori chapters, the image was considered
as the output of a linear shift invariant operator S to white noise, or, in a deterministic setting,
to a signal u(ej ) l2 , |u(ej )| = 12 . This leads to the set-up shown in the Figure 3.3., where
G(z) represents a nominal model of a particular texture, y and s denote the ideal and actual images, respectively, and where the (unknown) operator describes the mismatch between these
two images.
Under the proposed framework, apply the model (in)validation theorem 1 to the problem
set-up for the texture recognition. The problem will be converted to be the following form:
Find for G(z) an input u, |u(e(j) |=1 and an admissible uncertainty operator of minimum
size opt :
.
opt = min,u {|||| : s = [( + I)Gu] + z}
where ||.|| denotes some norm of interest.
Consider opt as the criteria for texture classification, where the small opt means the small
mismatch, and vice versa.
3.3.1
For texture recognition problem, the the image is considered as the output of a model which
has single input and multi output(SIMO). The texture recognition under this condition is a
(in)validation procedure for SIMO system.
Then we have:
q(ej ) = P11 (ej )p(ej ) + P12 (ej )u(ej )
s(ej ) = P21 (ej )p(ej ) + P22 (ej )u(ej ) + z(ej )
23
P11
0
0
. 0
0 0
0 0 0
P21
0
0
0
1
0
0
. 0
0 0
0 0 1
"
Maug
.
=
M11 M12
M21 M22
.
=
0 0
iT
P11 (ej )
P21 (ej )
0 0
1
j
s(e )
|
P12 (ej )Su (ej )
|
1 P22 (ej )Su (ej )
M11 , M12 , M21 , M22 can be expressed as M11 = a11 + jb11 , M12 = a12 + jb12 , M21 = a21 + jb21
and M22 = a22 + jb22 . Based on the discussion of the complex LMI matrix problem, we can get
the following results for general cases.
"
.
L(x) = M (ej )
X() 0
0
1
"
j
M (e )
X()
0
0
y()
< 0,
In the realization of LMI algorithm of Matlab, the function LMITERM has a requirement,
that the value can not be complex. Because of this reason, the forms should be adjusted for the
requirement of Matlab.
"
a b
b a
<0
L(x) < 0
Re{L(x)} Im{L(x)}
Im{L(x)} Re{L(x)}
Where
"
Re{L(x)} =
#
a011 Xa12 + b011 Xb12 (a021 a22 + b021 b22 )
a012 Xa12 + b012 Xb12 (a022 a22 + b022 b22 ) + y()
24
"
3.3.2
a011 Xb11
a012 Xb11
b011 Xa11
b012 Xa11
(a021 b21
(a022 b21
Im{L(x)} =
#
Examples
Example 1:
Assume G(z) =
2
z 2 +0.1z0.12
and (z) =
0.05
.
z 2 +0.1z0.12
Then we have:
0.0001
0.0056444
0.011189
0.016733
0.022278
0.027822
0.033367
0.038911
0.044456
0.05
opt
=0
2.1927
1.5666
1.0596
0.66545
0.37802
0.18754
0.08281
0.031768
0.0083003
0.00043026
0.011
=5
2.4214
1.7963
1.2918
0.8878
0.5725
0.33985
0.18215
0.086101
0.031474
0.0073614
0.014
= 20
2.2744
1.6498
1.1464
0.75612
0.47521
0.28768
0.16787
0.097067
0.055328
0.030491
0.012
= 30
2.2034
1.6142
1.1458
0.77667
0.49066
0.27957
0.13641
0.054631
0.016178
0.002356
0.012
= 55
2.4982
1.8203
1.2754
0.84857
0.52319
0.28768
0.13984
0.062837
0.024155
0.0054036
0.013
Example 2:
In the following Figure, apply the proposed algorithm to get opt and use it to do texture
classification.
25
0.04
0.06
0.08
0.1
opt
I 1,2 (
y)
1.4
0.94892
0.67078
0.48445
0.057
I 1,3 (
y)
1.5497
1.0636
0.77038
0.5856
0.061
I 1,4 (
y)
1.5769
1.1234
0.83756
0.64198
0.07
I 1,5 (
y)
49099
46615
44241
41988
0.94
I 1,6 (
y)
76394
71840
67455
63206
0.97
I 1,7 (
y)
41229
39354
37552
35835
0.92
I 1,8 (
y)
39230
37373
35606
33897
0.92
From the above results, we can see that, the opt can be a criteria for texture recognition.
Figure 3.4 is the figure 1 corrupted by additive noise with ||.||2 < 5. Figure 3.5 and 3.6 are parts
in different positions of the same type texture background as figure 1.
Example 3:
To test the algorithm more, we shift the area under the same background as Figure 3.5 with
different horizontal and vertical directions, and get different opt to check the result. Assume
the noise is additive with ||.||2 < 5, and the result is listed in Table 3.3, 3.4.
26
0.04
0.06
0.08
0.1
opt
0.04
0.06
0.08
0.1
opt
I 1,2 (
y)
0.088986
0.046072
0.027926
0.018476
0.01
1,8
I (
y)
1.7557
1.2309
0.89328
0.67204
0.065
I 1,3 (
y)
0.24264
0.1356
0.081752
0.053295
0.015
I 1,9 (
y)
2.0081
1.4193
1.0364
0.77355
0.078
I 1,4 (
y)
0.43166
0.27629
0.18808
0.13633
0.02
1,10
I
(
y)
2.373
1.6936
1.2492
0.93463
0.097
I 1,5 (
y)
0.6129
0.39362
0.267
0.19479
0.03
1,11
I
(
y)
2.7161
1.968
1.4599
1.1043
0.11
I 1,6 (
y)
0.98534
0.66173
0.46939
0.34955
0.038
I 1,12 (
y)
3.157
2.3733
1.7906
1.3607
0.12
I 1,7 (
y)
1.4239
0.97687
0.70071
0.52469
0.059
I 1,13 (
y)
3.7224
2.8596
2.2186
1.7468
0.13
0.04
0.06
0.08
0.1
opt
I 1,2 (
y)
14730
13882
13099
12359
0.95
I 1,3 (
y)
12512
11660
10870
10130
0.95
I 1,4 (
y)
13034
12250
11528
10852
0.952
I 1,5 (
y)
13123
12290
11527
10821
0.95
I 1,6 (
y)
13908
13026
12206
11446
0.95
I 1,7 (
y)
13425
12658
11945
11278
0.95
27
We can see that under vertical shift condition, the opt criteria works well. But under
horizontal shift condition, it didnt work because the system model is setup for vertical direction.
If we improved the theorem into 2-D direction and transform the problem into 2-D condition,
it may work better.
Example 4:
In these tests, the images have been made zero mean and normalized to ||.||2 = 1. The
images are as the Figure 3.6, and noise level was set to = 0.05, the results are listed in Table
3.5.
Table 3.5 and the results of I(
y)
0.04
0.06
0.08
0.1
opt
0.04
0.06
0.08
0.1
opt
I 1,2 (
y)
0.11848
0.088952
0.067472
0.052033
0.001
I 1,8 (
y)
1.173
1.0336
0.92049
0.82436
0.06
I 1,3 (
y)
0.26697
0.21973
0.18425
0.1566
0.01
1,9
I (
y)
1.3167
1.1727
1.0545
0.95299
0.08
I 1,4 (
y)
0.43647
0.37158
0.32074
0.27923
0.016
I 1,10 (
y)
1.5262
1.3635
1.2286
1.1137
0.11
I 1,5 (
y)
0.5769
0.49201
0.42736
0.37536
0.02
1,11
I
(
y)
1.633
1.4653
1.3232
1.2009
0.12
I 1,6 (
y)
0.77889
0.67288
0.58972
0.52213
0.03
1,12
I
(
y)
1.7965
1.6138
1.4596
1.326
0.13
I 1,7 (
y)
1.0027
0.87922
0.78091
0.69937
0.04
1,13
I
(
y)
2.017
1.8166
1.645
1.4959
0.15
The image size is 64 10. In Figure 3.6, the every vertical shift step shift the image with
5 pixels. From Table 3.5, we can see that the largest shift will be 30 pixels, when require
opt < 0.05
For the different type textures, the test sample texture images are as Figure 3.8, and the
invalidation result of opt was listed in Table 3.6.
Figure3.8
28
0.04
0.34333
0.64667
0.95
opt
3.4
I 1,2 (
y)
340.93
138.28
31.073
0.29715
0.9
I 1,3 (
y)
493.07
221.22
59.164
1.0318
0.95
I 1,4 (
y)
424.33
187.22
49.913
0.85273
0.94
I 1,5 (
y)
400.49
169.06
42.808
0.7315
0.94
I 1,6 (
y)
520.88
239.07
67.059
1.264
0.97
I 1,7 (
y)
364.44
152.12
36.17
0.5331
0.91
model of the image. The following part will apply the (in)validation algorithm based on the
2-D filters image model, and the results will show the efficiency of the model and the validation
algorithm.
3.4.1
Examples
Example 1:
The images are chosen in different position of the whole texture background. This image
has some periodic property with noise. We choose as the 5% energy of the input (||.||2 ). The
size of the image is 94 94. For model reduction, we choose lr = 3. It can be showed that for
lr = 3, the image can be reconstructed well. The results are shown in Figure 3.9 and Table 3.7.
Iv1,n (
y ) and Ih1,n (
y ) express the vertical and horizontal filter model respectively.
Table 3.7 and the results of I(
y)
0.01
0.105
0.2
Iv1,2 (
y)
0.036999
0.013094
0.006261
Ih1,2 (
y)
0.032742
0.016105
0.0093659
Iv1,3 (
y)
0.085412
0.035724
0.018607
Ih1,3 (
y)
0.09073
0.03841
0.01941
Iv1,4 (
y)
0.16259
0.082398
0.052301
Ih1,4 (
y)
0.11996
0.068086
0.047621
Iv1,5 (
y)
0.22965
0.098114
0.058963
Ih1,5 (
y)
0.2662
0.12419
0.091537
0.01
0.105
0.2
Iv1,6 (
y)
0.25676
0.12485
0.083035
y)
Ih1,6 (
0.34653
0.16924
0.12733
Iv1,7 (
y)
0.3588
0.17578
0.11544
y)
Ih1,7 (
0.44825
0.2484
0.18254
Iv1,8 (
y)
0.46859
0.2653
0.17821
y)
Ih1,8 (
0.8556
0.40442
0.23445
Iv1,9 (
y)
0.56763
0.23936
0.16509
y)
Ih1,9 (
0.6451
0.29128
0.19548
0.01
0.105
0.2
y)
Iv1,10 (
0.57209
0.21921
0.15502
Ih1,10 (
y)
0.62187
0.25599
0.15626
y)
Iv1,11 (
0.67865
0.25571
0.17614
Ih1,11 (
y)
0.79029
0.34455
0.1967
y)
Iv1,12 (
0.75834
0.25891
0.18127
Ih1,12 (
y)
0.61524
0.30155
0.20498
29
From table 3.7, we can see that all opt < 0.01. The results was better than takeing SVD
before. The reason may be that after SVD decomposition and model reduction, only key information of the texture structure remains.
Example 2:
This image has some periodic property with noise. We also choose as the 5% energy of the
input (||.||2 ). The size of the image is 64 64. For model reduction, we choose lr = 8. It can be
showed that for lr = 8, the image can be reconstructed well. The results are shown in Figure
3.10 and Table 3.8.
Table 3.8 and the results of I(
y)
0.01
0.105
0.2
Iv1,2 (
y)
0.055463
0.0072827
0.0021096
Ih1,2 (
y)
0.066322
0.0089614
0.0032282
Iv1,3 (
y)
0.29882
0.10738
0.059401
Ih1,3 (
y)
0.41178
0.12704
0.072031
Iv1,4 (
y)
0.43431
0.12664
0.069125
Ih1,4 (
y)
0.32709
0.10518
0.056889
Iv1,5 (
y)
0.47972
0.027096
0.012982
Ih1,5 (
y)
0.52293
0.042722
0.012937
0.01
0.105
0.2
Iv1,6 (
y)
0.51972
0.039081
0.019593
Ih1,6 (
y)
0.56625
0.058679
0.020106
Iv1,7 (
y)
0.83892
0.17797
0.10507
Ih1,7 (
y)
0.8488
0.12633
0.050167
Iv1,8 (
y)
1.3121
0.15372
0.083288
Ih1,8 (
y)
1.3122
0.23267
0.070809
Iv1,9 (
y)
1.3715
0.18303
0.10223
Ih1,9 (
y)
1.4422
0.22995
0.077837
0.01
0.105
0.2
Iv1,10 (
y)
1.6523
0.2423
0.10756
Ih1,10 (
y)
1.6452
0.27469
0.061891
Iv1,11 (
y)
2.1183
0.37516
0.11816
Ih1,11 (
y)
2.0547
0.42783
0.076685
Iv1,12 (
y)
1.6783
0.40574
0.19141
Ih1,12 (
y)
1.5033
0.35879
0.16596
30
From table 3.8, we can see that all opt < 0.1. These textures have the same type texture
construction.
Example 3:
For different type texture, the results are list below. For comparisons, we choose noise
= 10% energy of input. The sample images are as Figure 3.11. The results are listed in Table
3.9.
Table 3.9 and the results of I(
y)
0.1
0.55
0.9
opt
Iv1,2 (
y)
0.071633
0.012426
0.00019605
<0.01
y)
Ih1,2 (
0.060475
0.012054
0.00021195
<0.01
Iv1,3 (
y)
5.7204
0.89536
0.0019302
0.5
y)
Ih1,3 (
5.9574
0.9928
0.003285
0.5
Iv1,4 (
y)
18.75
5.2887
0.12652
0.70
y)
Ih1,4 (
21.018
5.6141
0.15009
0.75
Iv1,5 (
y)
28.058
7.7955
0.20527
0.85
y)
Ih1,5 (
24.288
6.977
0.3314
0.83
3.4.2
There were one problem worth mentioning during analyzing the experiment results. One
possible problem to get unexpected abnormal result is the difference of the +/- sign when taking
SVD decomposition. As we all know, the sign change of u and v during SVD decomposition will
not change the restoration result of the original Matrix.
I=
l
X
i=1
i ui viT
l
X
i=1
i (ui )(vi ) =
l
X
i ui , vi1 = i vi ,
i=1
Because of this, the validation result will be dramatically affected. Some examples were
tried to testify this assumption. The results are listed below.
Take I1 = [u1, u2, u3] R943 , consider 3 conditions:
I2 = [u1, u2, u3] R943 ,
I3 = [u1, u2, u3] R943 ,
I4 = [u1, u2, u3] R943 .
31
I 1,2 (
y ) I 1,3 (
y ) Iv1,4 (
0.1
0
254.32
0
0.4
0
113.16
0
0.7
0
28.307
0
opt
0
0.95
0
32
model images exhibiting a given texture as the realization of a second order stationary stochastic
process. It considers the intensity values I(k, :) of the k th row of the image as the output of a
discrete linear shift-invariant, not necessarily causal, system driven by white noise. Moreover,
proved in [46], finding texton and completing missing portions of a textured image reduce both
to a rank minimization problem.
The frame work will be introduced in the following, which is the base of my future work to
extend the current inpaintg work using Roesser models and the 2-D Hankel Matrix.
4.1
Texture modelling
In [46], the textured images with the intensity values I(k, :) of the k th row of the n m
image is modelled as the output, at step k, of a discrete linear shift-invariant system driven by
the white noise u:
n
X
I(k, :) =
i=1,j6=k
where ai , bi are the unknown parameters to be extracted from the image data.
The issue of causality can be addressed by considering the given n m image as one period
of an infinite 2-D signal with period (n,m). Thus it admits a state space realization of the form:
xk+1 = Axk + Buk
yk = Cxk , An = I
To extract the texture model parameters from noisy images, finite horizon model reduction
method (algorithm 1) is applied here, which is based on the SVD of a circulant Hankel matrix
constructed from the image data, where
.
HI =
R1
R2
..
.
Rn
R2 Rn
R3 R1
.. . .
..
.
.
.
R1 Rn1
RiT denote its ith row of the image matrix: I Rmn , form nm n block circulant Hankel
matrix HI .
4.2
Finding textons
Texton is the smallest sub-image to reproduces the original image. In [46], finding textons
33
problem can be solved by finding regions of the image corresponding to local minima of the rank
of the associated Hankel matrices. The main result was given as the following:
1. Given an nm image I(x, y). (A,B,C) denotes the state-space matrices of the corresponding model. Assume that the image I contains at least one complete texton, that is, there
exists some r < min{m, n}, such that Ri = CAi1 B, where A Rrr , Ar =I and Ak 6=I
for any 1 k < r.
For an ideal image, uncorrupted by noise, form the following matrix in this case:
Hnk
.
=
R1
R2
..
.
R2 Rnk
R3
R1
.. . .
..
.
.
.
R1 Rnk 1
Rnk
.
where nk = (N 1)r + k, satisfies rank(H(N 1)r+k )> r,
Textons can be found by considering the rank of a sequence of Hankel matrices, staring
with k = n, with decreasing values of k, and searching for relative minima of rank(Hk ).
2. For the more realistic case of ideal texture, corrupted by additive noise v, the Hankel
matrices of the actual (Yk ) and ideal (Hk ) images are related by
Yk = Hk + Hv
the equivalent problem is
(
(Hv ) = (Yk Hk )
Hk H , rank(Hk ) = r
Figure4.1
34
4.3
Texture Inpaiting
The texture inpainting problem arises in the the context of restoring a textured image where
some pixels. It can also be used to remove unwanted elements from the image or to correct
errors. As shown in [46], this problem can be recast into a rank minimization problem. The
main result was given as the following:
1. Given an n m image I(x, y), let H and (A,B,C) denote the associated Hankel matrix and
state-space model, respectively. Assume that the image I contains at least one complete
period, that is A Rrr , with Ar =I and r <min{m, n}.
Assume now that R1 , the first row of image, is missing or corrupted. The corresponding
Hankel matrix is given by:
x
R2
..
.
.
H(x) =
R1
..
.
Rr
x
..
.
R1
R2
.
..
. ..
x
.
..
. ..
Rr
..
.
R2
R1
..
.
Rr
.
..
R1 Rr1 Rr
(Y H(x))
H(x) H , rank(H(x)) = r
Y and H(x) denote the corresponding image and the underlying low rank Hankel operator.
Reducing the Computational Complexity:
A potential difficulty with the approach outlined above stems from the fact that rank minimization problems are known to be generically NP-hard. The computationally tractable convex
relaxations are given in [46] based on the specific structure of the problem.
Given the Hankel matrix H(x) with rank r, so does the Toeplitz matrix:
.
T (x) =
R1
R2
..
.
x
R1
..
.
Rn1
Rn1
x
..
..
.
.
R1
Proven in [46], one can attempt to solve the above NP-hard problem by solving the following
optimization problem:
min
log((i)2 + )
35
(Y H(x))
H(x) H , rank(H(x)) = r
Here we deal with the rank minimization problem with a patch wise optimization approach,
which is the most different from the pixel-wise methods indicated before. This image-based
approach is very quick compared to the other mentioned methods, and also very efficient to the
inpaiting problem. The small patches as a whole can also preserve the inherent property better
for the same type class of texture. The figure 4.2 gives the illustration of this.
36
viewpoint of control and operator theory, where images exhibiting a given texture are viewed
as the output of an unknown system with periodic impulse response, corrupted by noise. Many
problems of practical interest require addressing the issues of identification and model reduction
of systems having a periodic impulse response. Currently available techniques are not well suited
for solving these problems, since they cannot guarantee the key structural periodicity properties.
In the proposal, I introduced the the finite horizon model reduction method and extended
the method from 1-D to 2-D motivated by the idea of 2-D IIR filter realization and Roessers
2-D state-space model. Besides, many experiments have been done for the applications of semiblind (in)validation theory to image recognition and classification. The results have shown the
efficiency, drawbacks and future work orientation of the method. Finally, a Hankel operator
approach is applied in image restoration and inpaiting by converting to a rank minimization
problem. The discussion and experiments were done for both procedural-based and image-based
algorithm.
Future Work:
Based on the work done in the proposal, the following lists my recent future work for my
Ph.D degree with rough time schedule.
1. Aug. 05 Sep. 05
Summarization of finished work into papers, including the work on the applications of
semi-blind (in)validation theory to texture classification and the work on 2-D extension
of the Hankel operator texture modelling approach. There will be additional work on
theoretical or experimental analysis.
2. Sep. 05 Nov. 05
Consummating the current 2-D texture image processing frame work. The approximation
error should be analyzed. More efficient modelling method is still in need, such as using
a special single Hankel matrix to get system realization. There are much work needed to
extended many applications from 1-D to 2-D, like the image inpaiting.
3. Oct. 05 Dec. 05
Extending the current work to dynamic texture. As we can see in the proposal and other
related papers, the algorithms proposed can be modified to incorporate time evolution of
37
the state representation, potentially providing for efficient ways of modelling and recognizing dynamic texture. Dynamic texture analysis is getting considerable attention currently.
4. Nov. 05 Feb. 06
Generalizing the current semi-blind (in)validation results to cases involving time varying
and slowly time varying uncertainty structure. Developing necessary and sufficient invalidation conditions for mixed LTI/LTV and LTI/SLTV structures. Applications of the new
results to static and dynamic texture processing.
5. Mar. 06 Jun. 06
Other unscheduled works related to the research work and the thesis.
38
References
[1] L. F. Gool, P. Dewaele, and A. Oosterlinck, Survey: texture analysis anno, cvgip, vol. 29,
pp. 336-357, 1985
[2] M. Tuceryan and A. K. Jain, Texture analysis, in Handbook of Pattern Recognition and
Computer Vision,. H. Chen, L. F. Pau, and P. S. Wang, Eds. World Science Publishing,
1993, pp. 235-276.
[3] T. Ojala, M. Pietikainen, and D. Hardwook, A comparative study of texture measures
with classification based on feature distributions, Pattern Recognition, vol. 29, no. 1, pp.
51-59, 1996
[4] T. Randen and J. H. Husoy, Filtering for texture classifications: Acomparative study,
pami, vol. 21, no. 4, pp. 291-310, 1999
[5] J. Zhang and T. Tan, Brief review of invariant texture analysis methods, Pattern Recognition, vol. 35, pp. 735-747, 2002.
[6] B. Julesz, Visual pattern discrimination, IRE Trns of Information Theory, vol. IT, no. 8,
pp. 84-92, 1962
[7] G. R. Cross and A. K. Jain, Markov random field texture models, IEEE Trans. on Pattern
Analysis and Machine Intelligence, vol. 5, pp. 713-718, 1983
[8] J. Mao and A. K. Jain, Texture calssification and segmentation using multi-resolution
simultaneous autoregressive models, Pattern Recognition, vol. 25, pp. 173-188, 192
[9] R. Chellappa and R. Kashyap, Texture synthesis using 2d noncausal autoregressive models, IEEE Trans. on Acoustics, Speech and Signal Processing, vol. assp-33, no. 1, pp.
194-199, February 1985
[10] Y. Q. Xu, . Guo, and H. Shum, Chaos mosaic: Fast and memory efficient texture synthesis, Microsoft Research, Tech. Rep. MSR-TR-2000-32, April 2000.
[11] S. Geman and D. Geman, Stochastic relaxation, gibbs distributions and the bayesian
restoration of images, IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 6,
pp. 721-741, 1984
[12] D. Cano and T. H. Minh, Texture synthesis using hierarchical linear transforms, Signal
Processing, vol. 15, pp. 131-148, 1988
[13] M. Porat and Y. Y. Zeevi, Localized texture processing in vision: Analysis and synthesis
in gaborian space, IEEE Trans. Biomedical Engineering, vol. 36, no. 1, pp. 115-129, 1989
[14] J. D. Bonet and P. Viola, A non-parametric multi-scale statistical model for natural images. MIT Press, 1997, ch. 9
[15] D. Heeger and J. Bergen, Pyramid-based texture analysis/synthesis, in ACM SIGGRAPH, 1995
[16] K. Popat and R. W. Picard, Cluster-based probability model and itsapplication to image
and texture processing, IEEE Trans. on Image Processing, vol. 6, no. 2, pp. 268-284, 1997
[17] J. Portilla and E. P. Simoncelli, A parametric texture model based on joint statistics of
complex wavelet coefficients, Int. Journal of Computer Vision, vol. 40, no. 1, pp. 49-71,
2000
[18] A. Efros and T. Leung, Texture synthesis by non-parametric sampling, in ICCV, 1999
39
[38] Robert P. Roesser, A discrete state-space model for linear image processing, IEEE Transactions on Automatic Control, vol AC-20, no. 1, February 1975.
[39] Takao Hinamoto, Frederick W. Fairman, Separable-denominator state-space realization of
two-dimensional filters using a canonic form, IEEE Transactions on Acoustics, Speech,
and Signal Processing. vol ASSP-29, no. 4, Aug. 1981.
[40] B. L. Ho, R. E. Kalman, Effective construction of linear state variable models from inputoutput functions, Proc. 3rd Annu. Alberton Conf. Circuits and Syst. Theory, 1965, pp.
449-459.
[41] A. J. Tether, Construction of minimal linear state-variable models from finite input-output
data, IEEE Trans. Automat. Contr., vol. AC-15, pp.427-436, Aug. 1970
[42] P. Van Overschee and B. De Moor. Subaspce algorithm for the stochastic identification
problem, Automatica, vol. 29, no. 3, pp. 649-660, May 1993
[43] J. Chen and G. Gu, Control Oriented System Identification, An H Approach. New York:
John Wiley, 2000
[44] H. C. Lin, L. L. Wang, and S. N. Yang, Extracting periodicity of a regular texturebased
on autocorrelation functions, Pattern Recognition Letters, vol. 18, p. 433443, 1997
[45] S. Boyd, L. E. Ghaoui, E. Feron, and V, Balakrishnan, Lnear Matrix Inequalities in Systems
and Control Theory. Philadelphia: SIAM Studies in Applied Mathematics, 1994
[46] M. Sznaier,. A Hankel operator approach to texture modelling and inpainting.
41