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

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

74 4. A Variational approach for the <strong>de</strong>striping issue<br />

The previous PDE, also known as the heat equation, translates an isotropic diffusion process<br />

where all directions are smoothed i<strong>de</strong>ntically. The isotropic property of the heat<br />

equation is a major drawback for <strong>de</strong>noising applications. In fact, noise in homogeneous<br />

regions is effectively removed, however in heterogeneous areas, sharp discontinuities related<br />

to edges are processed similarly to noise. The corresponding high gradient magnitu<strong>de</strong><br />

is therefore reduced which results in a significant loss of contrast in the restored image.<br />

This limitation was first explored by [Perona and Malik, 1990] by means of anisotropic<br />

diffusion mo<strong>de</strong>l, where the intensity of the diffusion process is <strong>de</strong>pen<strong>de</strong>nt on the local<br />

gradient value. The <strong>de</strong>gree of smoothing is inversely proportional to the gradient value<br />

so that homogeneous areas are strongly diffused while edge-discontinuities are preserved.<br />

The anisotropic diffusion PDE proposed by Perona and Malik can be formulated as :<br />

∂u<br />

∂t (x, y, t) =div (φ(|∇u(x, y, t)|)∇u(x, y, t)) (4.4)<br />

where div is the divergence operator, ∇ <strong>de</strong>signates the spatial gradient operator and ψ is<br />

a <strong>de</strong>creasing function. The <strong>de</strong>creasing property of φ ensures that the smoothing process is<br />

weaker in regions with strong gradient values, and stronger in flat areas. Common choices<br />

for φ inclu<strong>de</strong> exponentially <strong>de</strong>creasing functions as :<br />

φ(|∇u|) =exp(− |∇u|2<br />

k 2 ) (4.5)<br />

If φ is a constant function, the PDE equation (4.4) of Perona and Malik is reduced to :<br />

∂u<br />

∂t (x, y, t) =div (∇u(x, y, t)) (4.6)<br />

which is another formulation of the isotropic heat equation. The anisotropic diffusion (4.4)<br />

also faces limitations. The presence of noise introduces strong oscillations which can be<br />

consi<strong>de</strong>red as edges and be subsenquently preserved. To circumvent this issue, alternative<br />

techniques were proposed in<strong>de</strong>pen<strong>de</strong>ntly in [Catté et al., 1992] and [Nitzberg and Shiota,<br />

1992]. The gradient of the image is replaced by a smoothed version and the PDE (4.4)<br />

becomes :<br />

∂u<br />

∂t (x, y, t) =div (φ(|∇(G σ ∗ u(x, y, t))|)∇u(x, y, t)) (4.7)<br />

where G σ is gaussian operator with variance σ. Further research in the field of anisotropic<br />

diffusion introduced in [Alvarez et al., 1992] resulted in a non linear PDE :<br />

∂u<br />

∂t (x, y, t) =g (|∇(G σ ∗ u(x, y, t))|) |∇u(x, y, t)| div<br />

( ∇u(x, y, t)<br />

|∇u(x, y, t)|<br />

)<br />

(4.8)<br />

where g is a <strong>de</strong>creasing function that tends to 0 when ∇u tends to infinity. The( second )<br />

<strong>de</strong>rivative of u in the direction orthogonal to the gra<strong>de</strong>nt ∇u is the term |∇u|div ∇u<br />

|∇u|<br />

,<br />

and coinci<strong>de</strong>s with the image level lines. The intensity of the diffusion process is regulated

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

Saved successfully!

Ooh no, something went wrong!