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.

91<br />

4.4 Destriping via gradient field integration<br />

This far, we introduced variational methods based on PDE’s and energy functional minimization,<br />

with image <strong>de</strong>noising applications in mind. These techniques are actually used<br />

extensively in computer vision/graphics, to solve over-constrained geometric problems.<br />

It is typically the case of Photometric Stereo Methods (PSM) and Shape From Shading<br />

(SFS) applications where the goal is to recover a <strong>de</strong>pth map (or an image) by integrating<br />

a gradient field with discrete values.<br />

Let us <strong>de</strong>note by G =(G x ,G y ) a bidimensional gradient vector <strong>de</strong>fined in a subspace<br />

Ω of R 2 . The problem of gradient field integration consists in <strong>de</strong>termining a function u<br />

whose gradient ∇u is close to G. Two classes of techniques can be used to tackle this<br />

ill-posed problem and estimate the true image u. Local integration techniques [Coleman<br />

et al., 1982], [Healey and Jain, 1984] and [Wu and Li, 1988] rely on a curve integration :<br />

∫<br />

u(x, y) =u(x 0 ,y 0 )+ G x dx + G y dy (4.69)<br />

where γ is the integration path from a starting point (x 0 ,y 0 ) to pixel (x, y) ∈ Ω. Local integration<br />

techniques recover an image u starting with an initial height and then propagating<br />

height values according to the neighborhood gradient values. Local integration techniques<br />

are <strong>de</strong>pen<strong>de</strong>nt on the data accuracy and can propagate error values when <strong>de</strong>aling with<br />

noisy gradient fields. Global integration methods [Horn and Brooks, 1986], [], [Frankot<br />

and Chellappa, 1988] and [Horn, 1990] can be formulated in a variational framework as<br />

the minimization of an energy functional :<br />

∫<br />

E(u) =<br />

Ω<br />

γ<br />

( ) ∂u 2 ( ) ∂u 2<br />

∂x − G x +<br />

∂y − G y dx dy (4.70)<br />

Unlike <strong>de</strong>noising and <strong>de</strong>composition variational mo<strong>de</strong>ls <strong>de</strong>scribed this far, both terms in the<br />

energy functional (4.70) are fi<strong>de</strong>lity terms that measure the L 2 norm difference between the<br />

gradient field components of u and the observed gradient field G. For instance, the energy<br />

functional does not inclu<strong>de</strong> any lagrange multiplier λ. The Euler-Lagrange equations of<br />

the previous functionals leads to :<br />

∂<br />

∂x<br />

( ∂u<br />

∂x − G x<br />

which can be simplified into the following Poisson equation :<br />

)<br />

+ ∂ ( ) ∂u<br />

∂x ∂y − G y = 0 (4.71)<br />

∇ 2 u = ∇.G (4.72)<br />

The previous equation can be solved using a fast marching method [Ho et al., 2006] or<br />

the streaming multigrid method proposed in [Kazhdan and Hoppe, 2008]. Among several<br />

other techniques, we draw a particular attention to the well-known Frankot-Chellapa

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

Saved successfully!

Ooh no, something went wrong!