Decision-based fuzzy image restoration for noise reduction based
on evidence theory
Tzu-Chao Lin
School of Department of Computer Science and Information Engineering, WuFeng University, Chiayi 62153, Taiwan, ROC
article info
Keywords:
Fuzzy theory
Evidence theory
Impulsive noise
Image restoration
abstract
A novel decision-based fuzzy averaging (DFA) filter consisting of a D–S (Dempster–Shafer) noise detector
and a two-pass noise filtering mechanism is presented in this paper. The proposed filter can effectively
deal with impulsive noise, and a mix of Gaussian and impulsive noise. Bodies of evidence are extracted,
and the basic belief assignment is developed using the simple support function, which avoids the coun-
ter-intuitive problem of Dempster’s combination rule. The combination belief value is the decision rule
for the D–S noise detector. A fuzzy averaging method, where the weights are constructed using a prede-
fined fuzzy set, is developed to achieve noise cancellation. A simple second-pass filter is employed to
improve the final filtering performance. Experimental results confirm the effectiveness of the new DFA
filter both in suppressing impulsive noise as well as a mix Gaussian and impulsive noise and in improving
perceived image quality.
Ó 2011 Elsevier Ltd. All rights reserved.
1. Introduction
Fuzzy theory and Dempster–Shafer (D–S) evidence theory deal
with fundamentally different types of uncertainty: vagueness and
ambiguity, respectively (Klir & Folger, 1998). They have been used
in numerous practical applications such as signal detection sys-
tems and image processing techniques (Bloch, 1996; Boston,
2000). Image processing techniques require a noise-free environ-
ment. However, images are often corrupted by impulsive noise
when obtained from a sensor or sent over a transmission channel.
Thus, preprocessing modules are a necessary part of image
processing such as edge detection, image segmentation, image re-
trieval, and image compression. The main goal of impulsive noise
reduction methods is to suppress noise while preserving fine de-
tails and edge elements. In this paper, we use D–S evidence theory
and fuzzy techniques to enhance the quality of images corrupted
by impulsive noise.
Compared to linear methods, nonlinear techniques have been
found to provide more satisfactory results due to their good perfor-
mance in noise removal (Astola & Kuosmanen, 1997). Many
nonlinear approaches have thus been developed for impulse noise
removal. For example, the median filter, as well as its modifications
and generalizations, has shown impressive performance in sup-
pressing impulsive noise (Ko & Lee, 1991; Lukac, 2002, 2004; Lukac
& Marchevsky, 2001; Lukac, Plataniotis, Smolka, & Venetsanopou-
los, 2004; Lukac, Smolka, Plataniotis, & Venetsanopoulos, 2006;
Yin, Yang, Gabbouj, & Neuvo, 1996). However, these approaches
are location-invariant in nature, meaning that they also tend to
modify some uncorrupted pixels. To avoid excessive smoothing
and to preserve image details, a number of switching median filters
with thresholding operations have been proposed (Abreu & Mitra,
1995; Chen, Ma, & Chen, 1999; Chen & Wu, 2001; Garnett, Huege-
rich, Chui, & He, 2005; Lin, 2008; Lin & Yu, 2004b; Smolka & Chyd-
zinski, 2005; Sun & Neuvo, 1994; Zhou & David, 1999). They
mainly use a decision-making process to separate noise-free pixels
from noisy ones. This way, noise-free pixels are left unchanged.
Since nonlinear filtering is activated only for noisy pixels, undue
distortion can be avoided. The typical decision rule is to employ
a single threshold for local signal statistics. Unfortunately, these
strategies may work well at a pre-assumed noise density level
but poorly at other noise density levels.
Arakawa and Lin et al. proposed an adaptive median scheme
controlled using fuzzy techniques and a partition fuzzy median fil-
ter based on the partition concept, respectively (Arakawa, 1996;
Lin & Yu, 2004a). In these cases, the filters are obtained as a
weighted sum of the input pixel and the output of the median
value. The optimized weights are obtained by training the specified
filters using a reference image. A good balance can be achieved be-
tween noise suppression and detail preservation. Although both of
these adaptive filters can provide satisfactory results, they highly
depend on the reference image and a relatively large number of
filter parameters. Another example of a decision-based filter is
Lin and Yu’s thresholding noise-free ordered mean filter based on
the D–S evidence theory (Lin & Yu, 2006). The filtering is achieved
by an efficient D–S impulse detector and a noise filter. Although
the D–S impulse detector depends on Dempster’s combination
0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2011.01.016
E-mail address: tclin@wfu.edu.tw
Expert Systems with Applications 38 (2011) 8303–8310
Contents lists available at ScienceDirect
Expert Systems with Applications
journal homepage: www.elsevier.com/locate/eswa