
Please cite this article in press as: J. Yuan, An improved variational model for denoising magnetic resonance images, Computers and Mathematics with
Applications (2018), https://doi.org/10.1016/j.camwa.2018.05.044.
J. Yuan / Computers and Mathematics with Applications ( ) – 3
2.2. Modified Bessel function of the first kind
The modified Bessel function of the first kind with order n is the solutions of classical Bessel’s modified differential
equations [18]. Its integral form can be formulated as follows:
I
n
(x) =
1
π
π
0
cos
(
nθ
)
exp
(
x cos θ
)
dθ, (4)
In addition, we review briefly the derivatives of the modified Bessel function I
0
(x), I
1
(x), that is,
˙
I
0
(x) = I
1
(x),
˙
I
1
(x) = I
0
(x) −
1
x
I
1
(x),
(5)
which will be used in ADMM algorithm.
2.3. Related works
According to Bayes’s rule, the MAP estimation of the noise-free image u can be worked out through the following:
¯
u = arg max
u
P
(
u|f
)
= arg max
u
P
(
f |u
)
P(u)
P(f )
= arg min
u
{
−log P
(
f |u
)
− log P(u)
}
(6)
In (6), let us utilize (3) to get
−log P
(
f |u
)
=
u
2
+ f
2
2σ
2
− log I
0
uf
σ
2
− log
f
σ
2
(7)
From [2], u satisfies a Gibbs prior, and can be represented as follows:
P(u) = exp
−γ
Ω
ϕ(u(x))dx
(8)
where ϕ denotes a nonnegative given function, γ is a constant. Ω is the considered image domain. In [15], ϕ(u(x)) = |∇u(x)|.
Hence, the MAP estimation model is
¯
u = arg min
u
u
2
+ f
2
2σ
2
− log I
0
uf
σ
2
− log
f
σ
2
+ γ ∥∇u∥
1
(9)
Based on the above model (9), in [15], authors add a convex data-fidelity term to this model. It can be rewritten as follows:
¯
u = arg min
u
u
2
+ f
2
2σ
2
− log I
0
uf
σ
2
− log
f
σ
2
+
1
σ
∥
√
u −
f ∥
2
2
+ γ ∥∇u∥
1
(10)
Moreover, authors have proved its existence and uniqueness of the solution. However, in this model, the gradient information
is not considered. Then, the method still brings about over-smoothed effects. Fig. 2 shows the over-smoothed effects via the
model in [15] for a manual image with Rician noise. From Fig. 2(c), it can be seen that the edges are over-smoothed.
3. Analysis of proposed model
3.1. Proposed model
Inspired by the model (10), a convex gradient data-fidelity term and a sparsity regularization are introduced into the MAP
estimation to get the following new denoising model:
¯
u = arg min
u
u
2
+ f
2
2σ
2
− log I
0
uf
σ
2
− log
f
σ
2
+
1
σ
∥∇u − ∇f ∥
2
2
+ γ ∥∇u∥
0
. (11)
For the above model (11), f is the given observed image, which is size of m × n. σ can be fixed previously. ∥∇u∥
0
denotes
the non-zero number of gradient of image. Since L
0
norm is a NP-hard problem, L
0
norm can be approximated to L
1
norm.
Therefore, the new denoising model can be rewritten as follows:
¯
u = arg min
u
u
2
+ f
2
2σ
2
− log I
0
uf
σ
2
+
1
σ
∥∇u − ∇f ∥
2
2
+ γ ∥∇u∥
1
, (12)
where u is the noise-free image. How to estimate σ will be discussed in Section 3.3. In (12), the convex gradient data-fidelity
term has the capability of retaining more details and cutting down over-smoothing. The sparsity regularization can reserve
main components of MR image and reduce the influence of Rician noise.