Summary: | 碩士 === 逢甲大學 === 電機工程研究所 === 86 === During the last few years, the nonlinear filter has been the
dominating filter class for removing the non-Gaussian noise.
The success of nonlinear filters isbased on two intrinsic
properties : edge preservation and effective noise attenuation
with robustness against impulse noise. However, it may cause
edge jitters, streaking and may remove important image details.
The main reasons is that the nonlinear filter use only rank-
order information of the input data within the filter window,
and discards the signal spatial-order information. In order to
utilize both rank- and spatial-order information of the input
data, several classes of nonlinear filter has been proposed.
Recently, Gibbs/MRF model has been widely used to model the
local statistics of images and proven to be a useful approach
for some image applications. The image enhancement technique,
introduced by Park and Kurz, utilize Gibbs/MRF model to
incorporate the spatial-order information in the filter
operation in order to reduce the noise effect of the image.
Compare with the conventional nonlinear filter, a better
performance for reducing the non-Gaussian noise of this approach
has been illustrated by some experimental results. However, due
to it*s computational complexity, it is difficult to enlarge the
filter*s window size in order to extract more spatial
information of the image data. In this research, we develop a
constrained Gibbs Markov Random Field (Gibbs/MRF) model for the
image processing. This approach utilizes the directional
constraint to choose more influential cliques for the Gibbs/MRF
model. In this way, we can not only reduce the computational
complexity, but also increase the window size of the filter. We
utilize the directional characteristic of the image feature to
include more influence cliques, and exclude the other cliques
for the Gibbs/MRF mode. By reducing the number of clique in the
Gibbs/GMF model, we can enlarge the window to acquire more
spatial-order information of the image without increasing the
computational complexity. The experimental results of proposed
method will also indicated in this thesis.
|