We refer the reader to the paper of
Lebrun et al. [17] for a complete description of the denoising
problem, as well as a detailed analysis and comparison of
state-of-the-art denoising methods. It has been shown by
Levin and Nadler [20] and Chaterjee and Milanfar [12] that
the current state-of-the-art denoising methods are close to
optimal when applied to natural images. Nonetheless, there is
still room for improvement in several directions. For instance,
while these methods manage to correctly remove most of
the noise, they tend to not properly recover some of the
image details. These methods also primarily deal with additive
Gaussian noise, whereas for many images the noise model
is unknown; in such cases, there is still ample room for
improvement (see Lebrun et al. [19] and references therein
for blind denoising algorithms). Our proposal in this paper is
to develop a strategy to improve any image denoising technique
by more carefully taking into account the local geometry
(direction of gradients and level-lines) of the image to
process.