rule, the optimized filters have demonstrated excellent robustness.
However, when conflicting evidence is present, Dempster’s combi-
nation rule often leads to counter-intuitive results (Zadeh, 1986).
In the present study, to avoid the counter-intuition problem and
to improve the quality of images corrupted by impulsive noise,
both data fusion and fuzzy techniques are incorporated into the
framework of our new decision-based filter. The new decision-
based fuzzy averaging (DFA) filter consists of two phases: the
detection phase and the filtering phase, which are based on D–S
evidence theory and fuzzy noise cancellation, respectively. An im-
pulse detection approach that uses the final belief values of the
classification of hypotheses to build the decision rule that distin-
guishes noise-free pixels from noise-corrupted pixels is proposed.
For noise cancellation, a fuzzy averaging mechanism, where the
weights are constructed using a predefined fuzzy set, is adopted.
Experimental results demonstrate that the proposed DFA filter
outperforms other decision-based filters in terms of both noise
suppression and detail preservation. Moreover, the proposed filter
provides satisfactory results in removing a mix of Gaussian and
impulsive noise.
The rest of this paper is organized as follows. Section 2 covers
the basic principles of D–S evidence theory. The design of the pro-
posed DFA filter is presented in Section 3. In Section 4, experimen-
tal results are provided to demonstrate the performance of the
proposed scheme. Finally, the conclusion is given in Section 5.
2. Dempster–Shafer evidence theory
The D–S evidence theory was first introduced by Dempster in
1960 and was later extended by Shafer (1976) and Guan and Bell
(1991). It allows the representation of both uncertainty and impre-
cision, and is generally recognized as a flexible alternative to the
Bayesian theory. This theory has inspired many researchers to
investigate the field of information fusion and related applications
(Bloch, 1996; Boston, 2000). Its goal is to decrease uncertainty in
information fusion by means of a combination rule applied to evi-
dence sources. The transferable belief model (TBM), which was
proposed and justified axiomatically by Smets, is different from
but closely related to the D–S evidence theory (Smets, 1990; Smets
& Kennes, 1994).
2.1. Basic concepts
Let
X
={h
1
, ..., h
n
} be a final set of possible hypotheses. This set
is referred to as the frame of discernment. The hypotheses in
X
are
assumed to be mutually exclusive and exhaustive. They can be
considered as not only single hypotheses (simple hypotheses),
but also as any union of hypotheses (compound hypotheses). The
power set is denoted by 2
X
. A basic belief assignment (i.e., mass
function m) is a function that assigns a value in [0, 1] to every sub-
set A of
X
and satisfies the following requirements:
mð/Þ¼0 ð1Þ
X
A # X
mðAÞ¼1 ð2Þ
where / is an empty set. A subset A with a nonzero mass m(A)>0is
called the focal element of m. The value of m(A) represents the belief
that is committed exclusively to A. Specifically, m(
X
) represents a
measure of ignorance. Focal elements and their masses construct
an evidence structure. In other words, a body of evidence is ex-
pressed in the form: {(A, m(A)) | A #
X
, m(A) > 0}. Notably, it is
not required that m(
X
)=1ormðAÞ 6 mðBÞ when A B.
The D–S evidence theory provides the representation of both
uncertainty and imprecision through the functions belief (Bel)
and plausibility (Pls), both of which are derived from the mass
function (m). Bel and Pls are respectively defined as:
Belð/Þ¼0
BelðAÞ¼
P
B # A
mðBÞ
8
<
:
ð3Þ
Plsð/Þ¼0
PlsðAÞ¼
P
B\A–/
mðBÞ
8
<
:
ð4Þ
where B # X and A # X. From (3) and (4), we derive the relationship
PlsðAÞ P BelðAÞ. The equation PlðAÞ¼1 Belð
AÞ holds, where
A ¼
X A is the complementary set of A. In the D–S evidence the-
ory, Bel(A) and Pls(A) may be regarded respectively as the minimum
and maximum uncertainty values about A. The belief interval
½BelðAÞ; PlsðAÞ provides a measurement of the imprecision about
the uncertainty value. Shafer showed that the evidence of any one
of the three functions m, Bel or Pls, is sufficient to derive the other
two (Guan & Bell, 1991)
mðAÞ¼
X
BA
ð1Þ
jABj
BelðBÞ; for all A # X ð5Þ
where |A B| denotes the cardinality number of the set (A B).
2.2. Dempster’s rule of combination
In addition to its capability in representing uncertainty of evi-
dence, the D–S evidence theory can also be used to combine bodies
of evidence. The theory provides a way of combining independent
bodies of evidence to increase confidence in the overall hypothesis.
The n combined bodies of evidence can be calculated by orthogonal
sum m ¼ m
1
m
2
m
n
for fusing independent information
sources m
i
. The orthogonal sum is associative and commutative
(Guan & Bell, 1991); it is defined in Dempster’s rule of
combination:
mð/Þ¼0
mðAÞ¼K
P
A
1
\\A
n
¼A
Q
n
i¼1
m
i
ðA
i
Þ
8
>
<
>
:
ð6Þ
where A # X and K
1
¼ 1
P
A
1
\...\A
n
¼/
Q
n
i¼1
m
i
ðA
i
Þ: Note that 1 K
1
can be interpreted as a measure of conflict among the bodies of evi-
dence to be combined. After the combination, a decision can be
made among the hypotheses according to the decision rule chosen.
There are three widely used decision rules: (a) maximum belief, (b)
maximum plausibility, and (c) maximum belief without overlap-
ping of belief interval (Hégarat-Mascle, Richard, & Ottlé, 2003).
The decision rules tend to be developed to fit the needs of specific
applications.
It is worth mentioning that combining conflicting evidence
using Dempster’s combination rule may produce counter-intuitive
results (Zadeh, 1986). To avoid some counter-intuitive effects that
may be encountered when combining contradictory bodies of evi-
dence, several authors have proposed other solutions (Smets,
1990). The K in (6) is a normalizing factor which intuitively mea-
sures how much the mass functions are conflicting. Under an
open-world assumption, Smets proposed his unnormalized combi-
nation rule (Smets, 1990):
mðAÞ¼
X
A
1
\\A
n
¼A
Y
n
i¼1
m
i
ðA
i
Þ; 8A # X ð7Þ
This rule implies that the mass of the null set may be nonzero,
which violates the condition expressed by (1); such cases reflect
contradiction in the state of belief. However, this cannot occur
8304 T.-C. Lin / Expert Systems with Applications 38 (2011) 8303–8310