08.02.2015 Views

Th`ese Marouan BOUALI - Sites personnels de TELECOM ParisTech

Th`ese Marouan BOUALI - Sites personnels de TELECOM ParisTech

Th`ese Marouan BOUALI - Sites personnels de TELECOM ParisTech

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

83<br />

Meyer’s work can then be used in hope of isolating the stripe noise from other structures<br />

present in the true scene.<br />

4.3.1 Yves Meyer’s mo<strong>de</strong>l for oscillatory functions<br />

The total variation regularization used in the previous section as a <strong>de</strong>noising technique<br />

can be seen from a different perspective as an image <strong>de</strong>composition mo<strong>de</strong>l, where<br />

the observed image f is approximated by a sketchy version u that lies in the BV space<br />

and the component v = f − u contains small sace <strong>de</strong>tails such as noise and/or texture.<br />

Many variational methods have the explicit goal of extracting an image u composed of<br />

homogeneous areas separated by sharp discontinuities but do not retain the component<br />

v as it is consi<strong>de</strong>red to be noise. This can be problematic for images containing texture<br />

because noise and texture are both oscillatory functions, processed equally with the TV<br />

regularization. From an image <strong>de</strong>composition perspective, ROF mo<strong>de</strong>l can be formulated<br />

as the following minimization :<br />

inf<br />

(u,v)∈BV (Ω)×L 2 (Ω)/u+v=f<br />

(<br />

)<br />

TV(u)+λ‖v‖ 2 L 2 (Ω)<br />

(4.43)<br />

In his investigation of the standard ROF mo<strong>de</strong>l [Meyer, 2002], Yves Meyer pointed out<br />

that small values of λ can remove fine <strong>de</strong>tails related to texture. To overcome this issue,<br />

he proposes a different <strong>de</strong>composition, where the classical L 2 norm associated with the<br />

residual v = f − u is replaced by a weaker-norm, more sensitive to oscillatory functions.<br />

Yves Meyer suggests finding a component v in the space G <strong>de</strong>fined as the Banach space<br />

composed of all distributions v which can be written as :<br />

v(x, y) =∂ x g 1 (x, y)+∂ y g 2 (x, y) (4.44)<br />

where g 1 and g 2 both belong to the space L ∞ (R 2 ). The space G is endowed with the norm<br />

‖v‖ G <strong>de</strong>fined as the lower bound of all L ∞ norms of the functions |⃗g| where ⃗g =(g 1 ,g 2 )<br />

and |⃗g(x, y)| = √ g 1 (x, y) 2 + g 2 (x, y) 2 . Additionnaly to G, which can be viewed as the<br />

dual space of BV , Meyer introduces the spaces E and F also suited to mo<strong>de</strong>l texture.<br />

The space E (dual of Ḃ 1,1<br />

1 ) is <strong>de</strong>fined similarly to G except that g 1,g 2 are in the space<br />

of boun<strong>de</strong>d mean oscillations functions <strong>de</strong>noted by BMO(R 2 ). For the space F (dual of<br />

H 1 ), g 1 ,g 2 belong to Besov space B∞<br />

−1,∞ (R 2 ). When the texture component v is assumed<br />

to lie in the space G, Yves Meyer proposes the following <strong>de</strong>composition mo<strong>de</strong>l :<br />

(<br />

TV(u)+λ‖v‖G(Ω ))<br />

2 (4.45)<br />

inf<br />

(u,v)∈BV (Ω 2 )×G(Ω 2 )/u+v=f<br />

The ‖.‖ G -norm can efficiently capture oscillating patterns because it is weaker than the<br />

‖.‖ 2 -norm (L 2 (Ω) ⊂ G(Ω) ). However, due to its mathematical form, the Euler-Lagrange<br />

equation of (4.45) cannot be expressed explicitly and several u + v <strong>de</strong>composition mo<strong>de</strong>ls<br />

have been later introduced as an approximation to Yves Meyer mo<strong>de</strong>l.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!