itimp09 (full)
Post on 13-Apr-2015
28 Views
Preview:
TRANSCRIPT
Fast cartoon + texture image filters
Abstract:
Can images be decomposed into the sum of a geometric part and a textural part. Ina
theoretical breakthrough, Yves Meyer [28] proposed variational models that force the
geometric part into the space of functions with bounded variation, and the textural part
into a space of oscillatory distributions. Meyer’s models are simple minimization
problems extending the famous total variation model. However, their numerical solution
has proved challenging. It is the object of a literature rich in variants and numerical
attempts.
This paper starts with the linear model, which reduces to a low-pass/high-pass
filter pair. A simple conversion of the linear filter pair into a nonlinear filter pair
involving the total variation is introduced. This new-proposed nonlinear filter pair retains
both the essential features of Meyer’s models and the simplicity and rapidity of the linear
model. It depends on only one transparent parameter: the texture scale, measured in pixel
mesh. Comparative experiments show a better and faster separation of cartoon from
texture. One application is illustrated: edge detection.
Introduction
The cartoon + texture algorithm decomposes any image into the sum of a cartoon
part, where only the image contrasted shapes appear, and a textural part with the
oscillating patterns. Such a decomposition is analogous to the classical signal processing
low pass-high pass filter decomposition. However, the cartoon part of an image actually
contains strong edges, and therefore all frequencies, up to the high ones, while a texture
can also contain middle and high frequencies. Thus, a linear decomposition is never clear
cut. It blurs out edges, and leaves behind some texture in the low pass filtered part.
To overcome this difficulty, the algorithm proposed, is a nonlinear low pass-high
pass filter pair. At each point, a decision is made of whether it belongs to the cartoon part
or to the textural part. This decision is made by computing a local total variation of the
image around the point, and comparing it to the local total variation after a low pass filter
has been applied.
Edge points in an image tend to have a slowly varying local total variation when
the image is convolved by a low pass filter. Textural points instead show a strong decay
of their local total variation by convolution with a low pass filter. The cartoon+texture
algorithm is based on this simple observation.
The cartoon part keeps the original image values at points termed as non-textural
points. At points identified as texture points, the cartoon part takes the filtered value. At
points where the cartoon/texture decision is ambiguous, a weighted average of them is
given. The texture part simply is the difference between the original image and its
cartoon part.
The presented algorithm deals with color images. The cartoon-texture
decomposition is made independently on all channels.
A grey level or color image will be denoted by f : (x,y) Є Ω → IR (respectively
IR3) where is an open subset of IR2, typically a rectangle or a square. An image f is
defined on a continuous domain by interpolating a digital image defined on a finite set of
pixels. We are interested in decomposing f into two components f = u + v, such that u
represents a cartoon or geometric (piecewise-smooth) component of f, while v represents
the oscillatory or textured component of f. The oscillatory part v should contain
essentially the noise and the texture.
Scope of the Project:
The proposed algorithm can be used in Multimedia Industry for fast generation of
cartoon images and textures.
Module Description:
As indicated in the algorithm, the cartoon+texture decomposition only requires
the application on the gradient image of two low-pass filters, which are performed
directly by a discrete convolution. The gradient is computed by the simplest centered
difference scheme. The main modules are:
Apply the low pass filter Lσ to the image f.
Compute the modulus of the image gradients of f and Lσ ∗ f.
Convolve these moduli with the Gaussian Gσ to get the local total variation of f
and Lσ ∗ f.
Deduce the value of λ(x) at each point in the image.
Deduce the value of the cartoon image as a weighted average of f and Lσ ∗ f.
Compute the texture as the difference u-f.
ALGORITHMS USED
The algorithm used is based on a theory proposed by Yves Meyer in “Oscillating
patterns in image processing and nonlinear evolution equations”.
DOMAIN
DIGITAL IMAGE PROCESSING
Digital image processing is the use of computer algorithms to perform image
processing on digital images. As a subfield of digital signal processing, digital image
processing has many advantages over analog image processing; it allows a much wider
range of algorithms to be applied to the input data, and can avoid problems such as the
build-up of noise and signal distortion during processing.
SOFTWARE REQUIREMENTS
MATLAB 7.0 AND ABOVE
MATLAB
MATLAB is a high-performance language for technical computing. It integrates
computation, visualization, and programming in an easy-to-use environment where
problems and solutions are expressed in familiar mathematical notation.
Typical uses include:
Math and computation
Algorithm development
Modeling, simulation, and prototyping
Data analysis, exploration, and visualization
Scientific and engineering graphics
Application development, including Graphical User Interface building
MATLAB is an interactive system whose basic data element is an array that does not
require dimensioning. This allows you to solve many technical computing problems,
especially those with matrix and vector formulations, in a fraction of the time it would
take to write a program in a scalar non-interactive language such as C or FORTRAN.
SYSTEM DESCRIPTION:
PRIOR WORK:
The general variational framework for decomposing f into u + v is given in Meyer’s models
as an energy minimization problem
Where F1, F2 ≥ 0 are functional and X1, X2are spaces of functions or distributions such
that F1(u) < ∞ and F2(v) < ∞ if and only if(u, v) ∈ X1 × X2. The constant λ > 0 is a
Tuning parameter. A good model for (1) is given by a choice of X1 and X2 so that if u is
cartoon And if v is texture, then F1(u) << F2(u) andF1(v) >> F2(v) (such conditions
would insure a clear cartoon +texture separation; in other words ,if u is only cartoon,
without texture, then texture Components must be penalized by F1, but not by F2, and
vice-versa).
The long story of this problem can be summarized in a list of proposed choices for both
spaces X1 and X2, and both functional F1(u) andF2(v). In fact the choice for F1(u) has
quickly converged to the total variation of u, that excludes strong oscillations but permits
sharp edges. The main point under discussion has been what spaceX2 would model the
oscillatory part. Since the discussion is complex, we refer to Table I and its legend, which
present the main models. This table extends the model classification outlined in, and
adopts the same terminology.
One of the first nonlinear cartoon + texture models is the Mumford and Shah model [30],
for image segmentation, where f ∈ L2() is decomposed into u ∈ SBV () ([12], [2], [29], ,
a piecewise-smooth function with its discontinuity set Ju included in a union of curves
whose overall length is finite, and v = f − u ∈ L2() represents the noise or the texture.
The minimization problem is
Where H1 denotes the 1-dimensional Hausdorff measure (the length if Ju is
sufficiently smooth), and λ > 0 is a tuning parameter. With the above notations, X1 =
SBV () is the De Giorgi space of special functions with bounded variation. F1is
composed of the first two terms in the energy from (3), while the third term is F2(v) = R
v2,the quadratic norm. It is difficult to solve this model in practice, because of its non-
convex nature coming from F1(u).An easier decomposition can be obtained by the
Rudin, Osher, and Fatemi (ROF) total variation(TV) minimization model [37] for image
de noising. Their functional is convex and therefore more amenable to efficient
minimization.
The variational model is
Denotes the total variation of u in , also denoted by T V (u) or by |u| BV (). The
component u belongs to the space of functions of bounded variation BV () = u ∈ L1() :
R |Du| <∞. This space penalizes oscillations (such as noise or texture), but allows for
piecewise-smooth functions, made of homogeneous regions with sharp boundaries. Since
almost all level lines (or isolines) of a BV function have finite length, the BV space is
considered adequate to model images containing shapes. These shapes can actually be
extracted by edge detection or by image binarization and morphology.
The bibliography on algorithms minimizing the ROF functional and its multi-
scale variants, Convex dual numerical methods have been tested Hybrid models with
wavelets are described in Models where the L2 norm is replaced by the L1 norm are now
classical.
In strong mathematical geometric arguments are put forward in favor of the T V -
L1 model: explicit solutions can be computed for simple geometric objects. These
examples demonstrate that, based on the perimeter/area ratio, shapes are unambiguously
put either in the TV part or in the L1 part. This study connects the T V -L1 model with
the classical morphological granulometry. Accurate regularity results for the level set
boundaries of minimize of the T V −L1 model is also given, in any dimension, in [1].
Probably the most popular T V minimization algorithm is Chambolle’s projection
algorithm [14]. Recent years have, however, shown a trend to abandon the BV norm and
replace it bya so-called “non-local” norm inspired from.
Yet, as pointed out in, T V −L2 or T V −L1do not characterize the oscillatory
components. Indeed, these components do not have small norms in Lp(), p ≥ 1, . To
overcome this drawback, Y. Meyer [28] proposed in his seminal book weaker norms to
replace k · k2L2 in the ROF model, that would better model oscillatory components with
zero mean. The Meyer model is
Where ||.||*_ is the norm in one of the following spaces, denoted by G, F or E
(defined below for (Ω = IR2).
Definition 1: A distribution v belongs to G if and only if v = div(~g) for some ~g ∈ (L1)2
in the distributional sense. The endowed norm is
The introduction of the spaces G, F and Eis motivated by the fact that highly
oscillatory signals or images have small norms in G, F or E. For instance, || cos nx||G =
1n. The presence of a non-BV part in images is corroborated by the experimental-
numerical study . However, the three norms proposed by Meyer are not expressed as
integrals and are therefore difficult to compute.
It is also difficult to set up the right value of λ for real images. This problem is
addressed. The numerical experiments have shown promising results and justified further
inquiries. There has been an extensive line of papers modifying and interpreting Meyer’s
models, and proposing minimization schemes: An extensive mathematical analysis of
Meyer’s model in a bounded domain is performed .For many formal properties of the G-
norm the reader can refer. In the G-norm is replaced by the H−1 norm. This approach
using Sobolev spaces with negative exponents was extending ding. The F = div (BMO)
variant was numerically studied. There have also been extensions intending to decompose
u into three components, namely BV , texture, and a residual (e.g., noise). In the model
(where the space G = ˙W −1,1 is approximated by Gp = ˙W −1,p for large p), this is done
by solving
In the norm ||v||G of v = div~g is approximated by ||pg1 + g22||p, p ≥ 1 which is
of course far from the real problem with p = ∞. Aujol etal. Addressed the original Meyer
problem and proposed an alternate method to minimize
Where the above mentioned variants and others are summarized. Following this
paper’s terminology, the funding models that inspired this line of research are T V − L2
(ROF) and the original Meyer models T V −div (L1), T V −div (BMO) (numerically tried
in. A simpler variant is T V −H−1, since also the H−1 norm is small on oscillatory
signals. The hierarchy of the spaces used for the oscillatory part is complex: div (L1) and
div (BMO) are distributional first derivatives of vector fields in L1 and BMO
respectively .The Besov model takes the oscillatory part v into which is a space of second
derivatives of functions satisfying a Zygmund regularity condition. Since this condition is
close to assuming a Lipschitz bound on the functions, it is fair to say that the Besov
model
Defines distributions that are second derivatives of functions that have (almost)
bounded gradients.
In conclusion (as also pointed out by Y. Meyer), the four spaces G = div(L1),
H−1 =_(H1), F = div(BMO) and can be considered as variants of each other, since they
all appear as first derivatives of (bounded like)functions. Experimental evidence does not
favor one of them.
Generalizing T V −H−1, a generic T V -Hilbertmodel [11] can be defined using a
smoothing kernel K. The associated Meyer energy is
This model has also been proposed in [19]. TheL2 norm of K ∗ (f − u) can be
substituted by an Lp norm, p ≥ 1. One obtains slightly better results with p = 1 [18]. Our
numerical trials yield no significant difference between T V -Hilbert and the other
mentioned T V − X models. Because of its simplicity, we shall retain this version in the
experiments after fixing adequately the kernel K. This is precisely the object of the next
section. The main goal of the manuscript is to propose here a simpler and faster model
than the variational model (5), while better separating cartoon from texture We wish to
recall here the function spaces notations used in the next sections. H0 = L2 denotes the
space o square-integrable functions The Sobolev space H1 is defined by
We will also make use of the space defined in the Fourier domain by the set of
functions and distributions
we formulate the linear cartoon +texture H1−H−1 model inspired from Y. Meyer,
which can be easily and rapidly solved in the Fourier domain in one step. Since this
model introduces blurring in the cartoon component u, we propose in a novel nonlinear
cartoon + texture model that retains the simplicity and efficiency of the linear one, while
the cartoon component u is piecewise-smooth and with sharp edges. illustrates numerical
comparisons between the linear model, the non linear minimization model (5) and the
proposed fast non linear model; an application to edge detection is also shown, together
with a discussion on the local texture scale.
II. LINEAR VERSION OF MEYER’S MODEL
In view of the multiplicity and complexity of nonlinear models, it seems
reasonable to first fix as a reference the best linear model. Separation of scales in images
is classically obtained by applying a complementary pair of low-pass and high-pass filters
to the data f, namely u =LPF(f), and then v = f − u = HPF(f). The T V −H1 model is
easily linearized by replacing the total variation R |Du| by the Dirichlet integral R |Du|2.
Then the most natural variation linear model associated with Meyer’s ideas isH1 − H−1.
Indeed, H−1 is dual to H1, in the same way as G is dual to BV. The low pass filter f → u
is obtained by the minimization
The meaning of σ4 will be shortly explained. This model can be compared with the
classical Tikhonov quadratic H1 − L2 minimization
Which is equivalent in the Fourier domain to the low-pass filter ˆu = 11+(2πσ|ξ|)2
ˆ f. This Wiener filter is known to remove high-frequency components due to the edges of
f, and not only those due to oscillations (See Fig. 1).Using the Fourier transform in (6),
the H1semi-norm of u is R |Du|2 = R(2π|ξ|)2|ˆu(ξ)|2and the H−1 semi-norm of v is R |
ˆv(ξ)|2(2π|ξ|)2 . This implies in particular that u − f = v has zero mean, since feasible
solutions satisfy ˆv(0) =0.Minimizing this quadratic functional (6) in yields in Fourier the
unique solution ˆu = ˆLσ ˆ f, where
The meaning of the parameter σ is now easily explained: if the frequency ξ is
significantly smaller than 12πσ, then the ξ frequency is kept in u, while if ξ is
significantly larger than 12πσ ,then the frequency ξ is considered a textural frequency and
attributed to v. Thus, the solution (u, v) = (Lσ ∗ f, (Id − Lσ) ∗ f) is nothing but a pair of
complementary low pass and high pass filters. Note that as σ → 0, Lσ → Id. We will also
consider the filter Kσ, where ˆKσ(ξ) =e−(2πσ|ξ|)4 , which behaves still more like the
characteristic function of the ball centered at zero with radius 12πσ .
It is worth mentioning that related linear and nonlinear three-term decompositions
f = u+v+w based on the H1−H−1 duality were introduced the linear case is the (H1,H0,
H−1) decomposition, while the nonlinear decomposition uses piecewise (H1,H0,H−1)
(where piecewiseH1 is the SBV space for the cartoon, combined with piecewise H−1 for
the texture).
The observed efficiency of the linear pair (Lσ, Id − Lσ) (see Figure 3) leads to
consider nonlinear versions that would retain its main feature, namely the excellent
extraction of the texture by a high pass filter Id−Lσ. On the other hand, the non-
oscillatory parts of the initial image f should be kept unaltered even if they have sharp
edges. This is of course impossible with a linear filter. Thus, a local indicator must be
built to decide at each point x whether it belongs to a textural region or to a cartoon
region. The main character of a cartoon region is that its total variation does not decrease
by low pass filtering. The main character of a textured region is its high total variation
due to its oscillations. This total variation decreases very fast under low-pass filtering.
Formalizing these remarks leads to define the local total variation (LTV) at x,
(note that Lσ can be substituted by Kσ). The relative reduction rate of LTV is defined by
afunction x 7→ λσ(x).
PROPOSED METHOD:
The observed efficiency of the linear pair (Lσ, Id − Lσ) (see Figure 3) leads to
consider nonlinear versions that would retain its main feature, namely the excellent
extraction of the texture by a high pass filter Id−Lσ. On the other hand, the non-
oscillatory parts of the initial image f should be kept unaltered even if they have sharp
edges. This is of course impossible with a linear filter. Thus, a local indicator must be
built to decide at each point x whether it belongs to a textural region or to a cartoon
region. The main character of a cartoon region is that its total variation does not decrease
by low pass filtering. The main character of a textured region is its high total variation
due to its oscillations. This total variation decreases very fast under low -pass filtering.
Formalizing these remarks leads to define the local total variation (LTV) at x,
(note that Lσ can be substituted by Kσ). The relative reduction rate of LTV is defined by
a function x 7→ λσ(x), given by
which gives us the local oscillatory behavior of the function f. If λσ is close to 0,
we have
which means that there is little relative reduction of the local total variation by the
low pass filter. If instead λσ is close to 1, the reduction is important, which means that the
considered point belongs to a textured region. Thus, a fast nonlinear low pass and high
pass filter pair can be computed by weighted averages of f and Lσ ∗ f depending on the
relative reduction of LT V . We can set
where w(x) : [0, 1] → [0, 1] is an increasing function that is constant and equal to
zero near zero and constant and equal to 1 near 1. In all experiments the soft threshold
function w is defined by
fixed to 0.25 and 0.5. If λσ(x) is small, the function f is non-oscillatory around x and
therefore the function is BV (or cartoon) around x. Thus u(x) = f(x) is the right choice. If
instead λσ(x) is large, the function f is locally oscillatory around x and locally replaced
by (Lσ∗f)(x). The choice of λσ = 12 as underlying hard threshold is conservative: it
permits to keep all step edges on the cartoon side, but puts all fine structures on the
texture side, as soon as they oscillate more than once. Of course changes in the
parameters a1 or a2 would slightly modify the separation results.
Since it is desirable to have a one-parameter method, it seems advisable to fix the
threshold function w once and for all, as has been done in all experiments. In that way the
method keeps the scale σ of the texture as the only method parameter. That this last
parameter cannot be avoided is obvious: textural details become shapes when their sizes
grow, and therefore should be moved from the texture to the BV side.
Advantages:
Edges are highly preserved.
It differentiates the sharp and flat regions.
The output appears visually to a cartoon image.
Applications:
The developed project can be widely used in multimedia industry.
Literature Survey:
Various image filters used in the previous papers are reviewed. And the filters
used in those papers are listed below.
[1]L. Rudin, S. Osher, and E. Fatemi. Nonlinear total variation based noise removal
algorithms. Physica D, 60(1-4):259–268, 1992.
A constrained optimization type of numerical algorithm for removing noise from
images is presented. The total variation of the image is minimized subject to constraints
involving the statistics of the noise. The constraints are imposed using Lanrange
multipliers. The solution is obtained using the gradient-projection method. This amounts
to solving a time dependent partial differential equation on a manifold determined by the
constraints. As t → ∞ the solution converges to a steady state which is the de-noised
image. The numerical algorithm is simple and relatively fast. The results appear to be
state-of-the-art for very noisy images. The method is noninvasive, yielding sharp edges in
the image. The technique could be interpreted as a first step of moving each level set of
the image normal to itself with velocity equal to the curvature of the level set divided by
the magnitude of the gradient of the image, and a second step which projects the image
back onto the constraint set.
∫ |Du| + ∫ |v|2
[2]T.F. Chan and S. Esedoglu. Aspects of total variation regularized L1 function
approximation. SIAM J. on Appl. Maths., 65(5):1817–1837, 2005.
The total variation based image denoising model of Rudin, Osher, and Fatemi has
been generalized and modified in many ways in the literature; one of these modifications
is to use the L1 norm as the fidelity term. We study the interesting consequences of this
modification, especially from the point of view of geometric properties of its solutions. It
turns out to have interesting new implications for data driven scale selection and
multiscale image decomposition.
∫ |Du| + ∫ |v|
[1] AND [2] contains the classic BV or SBV + Noise Models.
[3]Y. Meyer. Oscillating patterns in image processing and nonlinear evolution
equations. American Mathematical Society Providence, RI, 2001, proposes the filter
This paper proposes an adaptive color feature extraction scheme by considering
the color distribution of an image. Based on the binary quaternion-moment-preserving
(BQMP) thresholding technique, the proposed extraction methods, fixed cardinality (FC)
and variable cardinality (VC), are able to extract color features by preserving the color
distribution of an image up to the third moment and to substantially reduce the distortion
incurred in the extraction process. In addition to utilizing the earth mover's distance
(EMD) as the distance measure of our color features, we also devise an efficient and
effective distance measure, comparing histograms by clustering (CHIC). Moreover, the
efficient implementation of our extraction methods is explored. With slight modification
of the BQMP algorithm, our extraction methods are equipped with the capability of
exploiting the concurrent property of hardware implementation. The experimental results
show that our hardware implementation can achieve approximately a second order of
magnitude improvement over the software implementation. It is noted that minimizing
the distortion incurred in the extraction process can enhance the accuracy of the
subsequent various image applications, and we evaluate the meaningfulness of the new
extraction methods by the application to content-based image retrieval (CBIR). Our
experimental results show that the proposed extraction methods can enhance the average
retrieval precision rate by a factor of 25% over that of a traditional color feature
extraction method.
∫ |Du| + ∫ ||v||Ẁ-1,∞
[4]S.J. Osher, A. Sol´e, and L.A. Vese. Image decomposition and restoration using
total variation minimization and the H−1 norm. Multiscale Modeling and
Simulation, 1(3):349–370, 2003.
In this paper, we propose a new model for image restoration and image
decomposition into cartoon and texture, based on the total variation minimization
ofRudin, Osher, and Fatemi [Phys. D, 60 (1992), pp. 259–268], and on oscillatory
functions, which follows results of Meyer [Oscillating Patterns in Image Processing and
Nonlinear Evolution Equations, Univ. Lecture Ser. 22, AMS, Providence, RI, 2002]. This
paper also continues the ideas introduced by the authors in a previous work on image
decomposition models into cartoon and texture [L. Vese and S. Osher, J. Sci. Comput., to
appear]. Indeed, by an alternative formulation, an initial image f is decomposed here into
a cartoon part u and a texture or noise part v. The u component is modeled by a function
of bounded variation, while the v component is modeled by an oscillatory function,
bounded in the norm dual to | · | H1 0. After some transformation, the resulting PDE is of
fourth order, envolving the Laplacian of the curvature of level lines.
∫ |Du| + ∫ ||v||Ĥ-1
[5]T.M. Le and L.A. Vese. Image decomposition using total variation and
div(BMO). SIAM Multiscale Modeling and Simulation, 4(2):390–423, 2005.
This paper is devoted to the decomposition of an image f into u + v, with u a
piecewise-smooth or “cartoon” component, and v an oscillatory component (texture or
noise), in a variational approach. Meyer [Oscillating Patterns in Image Processing and
Nonlinear Evolution Equations, Univ. Lecture Ser. 22, AMS, Providence, RI, 2001]
proposed refinements of the total variation model (Rudin, Osher, and Fatemi [Phys. D, 60
(1992), pp. 259–268]) that better represent the oscillatory part v: the spaces of
generalized functions G = div(L∞) and F = div(BMO) (this last space arises in the study
of Navier–Stokes equations; see Koch and Tataru [Adv. Math., 157 (2001), pp. 22–35])
have been proposed to model v, instead of the standard L2 space, while keeping u a
function of bounded variation. Mumford and Gidas [Quart. Appl. Math., 59 (2001), pp.
85–111] also show that natural images can be seen as samples of scale-invariant
probability distributions that are supported on distributions only and not on sets of
functions. However, there is no simple solution to obtain in practice such decompositions
f = u+v when working with G or F. In earlier works [L. Vese and S. Osher, J. Sci.
Comput., 19 (2003), pp. 553–572], [L. A. Vese and S. J. Osher, J. Math. Imaging Vision,
20 (2004), pp. 7–18], [S. Osher, A. Sol´e, and L. Vese, Multiscale Model. Simul., 1
(2003), pp. 349–370], the authors have proposed approximations to the (BV,G)
decomposition model, where the L∞ space has been substituted by Lp, 1 ≤ p < ∞. In the
present paper, we introduce energy minimization models to compute (BV,F)
decompositions, and as a by-product we also introduce a simple model to realize the
(BV,G) decomposition. In particular, we investigate several methods for the computation
of the BMO norm of a function in practice. Theoretical, experimental results and
comparisons to validate the proposed new methods are presented.
[6]J.B. Garnett, P.W. Jones, T.M. Le, and L.A. Vese. Modeling oscillatory
components with the homogeneous spaces BMO−_ and W−_,p. UCLA CAM Report
07- 21, to appear in PAMQ, 2007.
This paper is devoted to the decomposition of an image f into u + v, with u a
piecewise-smooth or “cartoon” component, and v an oscillatory component (texture or
noise), in a variation approach. The cartoon component u is modeled by a function of
bounded variation, while v, usually represented by a square integrable function, is now
being modeled by a more refined and weaker texture norm, as a distribution.
Generalizing the idea of Y. Meyer [32], where v 2 F = div (BMO) = BM˙ O −1, we
model here the texture component by the action of the Riesz potentials on v that belongs
to BMO or to Lp. In an earlier work [26], the authors proposed energy minimization
models to approximate (BV, F) decompositions explicitly expressing the texture as
divergence of vector fields in BMO. In this paper, we consider an equivalent more
isotropic norm of the space F in terms of the Riesz potentials, and study models where
the Riesz potentials of oscillatory components belong to BMO or to Lp, 1 _ p < 1 (thus
we consider oscillatory components in BM˙ O _ or in ˙W _,p, with _ < 0). Theoretical,
experimental results and comparisons to validate the proposed methods are presented.
[5] & [6] proposes the following filter equation.
∫ |Du| + ∫ ||v||BMO-1
And introduces a key new feature: The norm of the oscillatory part v decreases
when v oscillates more. This is obtained by putting a norm on v that is actually a norm on
a primitive of v. The TV-H−1, TV-DIV(BMO) AND TV-Besov Models follow the same
pattern.
[7]J.B. Garnett, T.M. Le, Y. Meyer, and L.A. Vese. Image decompositions using
bounded variation and generalized homogeneous Besov spaces. Applied and
Computational Harmonic Analysis, 23(1):25–56, 2007 (UCLA CAM Report 05-57,
2005).
This paper is devoted to the decomposition of an image f into u + v , with u a
piecewise-smooth or “cartoon” component, and v an oscillatory component (texture or
noise), in a variational approach. Y. Meyer [Y. Meyer, Oscillating Patterns in Image
Processing and Nonlinear Evolution Equations, University Lecture Series, vol. 22, Amer.
Math. Soc., Providence, RI, 2001] proposed refinements of the total variation model [L.
Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algorithms,
Phys. D 60 (1992) 259–268] that better represent the oscillatory part v: the weaker spaces
of generalized functions G =div( L ∞ ) , F =div( BMO ) , and have been proposed to
model v, instead of the standard L 2 space, while keeping u BV, a function of bounded
variation. Such new models separate better geometric structures from oscillatory
structures, but it is difficult to realize them in practice. D. Mumford and B. Gidas [D.
Mumford, B. Gidas, Stochastic models for generic images, Quart. Appl. Math. 59 (1)
(2001) 85–111] also show that natural images can be seen as samples of scale invariant
probability distributions that are supported on distributions only, and not on sets of
functions. In this paper, we consider and generalize Meyer's ( BV , E ) model, using the
homogeneous Besov spaces , −2< α <0 , 1 p , q ∞, to represent the oscillatory part v.
Theoretical, experimental results and comparisons to validate the proposed methods are
presented.
[8]J.F. Aujol, G. Gilboa, T. Chan, and S. Osher. Structure texture image
decomposition - modeling, algorithms and parameter selection. International
Journal of Computer Vision, 67(1):111–136, 2006 (UCLA CAM Report 05- 10,
2005).
This paper explores various aspects of the image decomposition problem using
modern variational techniques. We aim at splitting an original image f into two
components u and ¿, where u holds the geometrical information and ¿ holds the textural
information. The focus of this paper is to study different energy terms and functional
spaces that suit various types of textures. Our modeling uses the total-variation energy for
extracting the structural part and one of four of the following norms for the textural part:
L2, G, L1 and a new tunable norm, suggested here for the first time, based on Gabor
functions. Apart from the broad perspective and our suggestions when each model should
be used, the paper contains three specific novelties: first we show that the correlation
graph between u and ¿ may serve as an efficient tool to select the splitting parameter,
second we propose a new fast algorithm to solve the TV ¿ L1 minimization problem, and
third we introduce the theory and design tools for the TV-Gabor model.
[7] & [8] proposes the filter equation as given below.
∫ |Du| + ∫ |K * v|2
Here, the main fact is that the model, H1 − H−1, boils down to the decomposition
into a classic low-pass and high-pass decomposition.
Reference and Bibliography:
W.K. Allard. Total Variation Regularization for Image Denoising, I. Geometric
Theory. SIAM Journal on Mathematical Analysis, 39(4):1150–1190, 2008.
L. Ambrosio. Variational problems in SBV and image segmentation. Acta
Applicandae Mathematicae,
17(1):1–40, 1989.
L. Ambrosio, N. Fusco, and D. Pallara. Functions of bounded variation and free
discontinuity problems. Oxford University Press, 2000.
F. Andreu-Vaillo, V. Caselles, and J.M. Maz´on. Parabolic Quasilinear Equations
Minimizing Linear Growth Functionals. Birkhauser, 2004.
G. Aubert and J.F. Aujol. Modeling very oscillating signals. Application to image
processing. Applied Mathematics and Optimization, 51(2):163–182, 2005.
G. Aubert and L. Vese. A variational method in image recovery. SIAM J. on
Num. Anal., 34:1948–1979, 1997.
J.F. Aujol, G. Aubert, L. Blanc-F´eraud, and A. Chambolle.Image decomposition.
Application to SAR images.Lecture Notes in Computer Science, 2695:297–312, 2003.
J.F. Aujol, G. Aubert, L. Blanc-F´eraud, and A. Chambolle. Image
Decomposition into a Bounded Variation Component and an Oscillating Component.
Journal of Mathematical Imaging and Vision, 22(1):71–88, 2005.
J.F. Aujol and A. Chambolle. Dual norms and image decomposition models.
International Journal of ComputerVision, 63(1):85–104, 2005.
J.F. Aujol and G. Gilboa. Constrained and SNR-based solutions for TV-Hilbert
space image denoising. Journal of Mathematical Imaging and Vision, 26(1-2):217–237,
2006.
J.F. Aujol, G. Gilboa, T. Chan, and S. Osher. Structuretexture image
decomposition - modeling, algorithms and parameter selection. International Journal of
Computer Vision, 67(1):111–136, 2006 (UCLA CAM Report 05-10, 2005).
A. Blake and A. Zisserman. Visual Reconstruction. The MIT Press Cambridge,
Massachusetts, 1987.
A. Buades, B. Coll, and J. M. Morel. A review of image denoising algorithms,
with a new one. Multiscale Modeling Simulation, 4(2):490–530, 2005.
A. Chambolle. An algorithm for total variation minimization and applications.
Journal of Mathematical Imaging and Vision, 20(1):89–97, 2004.
A. Chambolle and P.L. Lions. Image recovery via total variation minimization
and related problems. Numerische Mathematik, 76(2):167–188, 1997.
T.F. Chan and S. Esedoglu. Aspects of total variation regularized L1 function
approximation. SIAM J. on Appl. Maths., 65(5):1817–1837, 2005.
V. Duval, J.F. Aujol, and Y. Gousseau. The TVL1 model: a geometric point of
view. SIAM Journal on Multiscale Modeling and Simulation, 8(1):154–189, 2009.
top related