Fast Hessian Frobenius Norm Based Image Restoration
Pengfei Liu
School of Computer Science and Engineering,
Nanjing University of Science and Technology,
Nanjing, China
e-mail: liupengfei199091@163.com
Liang Xiao
School of Computer Science and Engineering,
Nanjing University of Science and Technology,
Nanjing, China
e-mail: xiaoliang@mail.njust.edu.cn
Abstract—A projected gradient algorithm (PGA) which is
derived from the majorization-minimization (MM) framework
has been proposed recently for Hessian-matrix Frobenius
norm regularization image restoration model so that it
currently provides state-of-the-art performance. Outside the
MM framework and for the sake of further accelerating the
convergence speed, this paper presents an efficient algorithm
for image restoration under the Hessian-matrix Frobenius
norm regularization. Using variable splitting to obtain an
equivalent constrained optimization formulation, then our
algorithm is addressed with an augmented Lagrangian
method. Under the alternating direction method of multipliers
(ADMM) framework, a fast algorithm with split augmented
Lagrangian shrinkage scheme is thus proposed for image
restoration. Finally, experimental results demonstrate that our
algorithm achieves better results than PGA in terms of peak
signal to noise ratio (PSNR) and convergence rate.
Keywords-Hessian Frobenius norm; image restoration;
alternating direction method; projected gradient algorithm
I. INTRODUCTION
Recently, image restoration that aims to reconstruct the
desired image from its blurry and noisy observation plays an
important part in various application areas such as medical
imaging, astronomy and image coding.
Assuming an image
ℜ→ℜ⊂Ω
2
:f as a function to
be continuously differentiable, and let
denotes the
observed image with spatial-invariant blur and additive
noise corrupted, thus the degradation model can be linearly
formulated as
efAg += , (1)
where A is a linear blur operator and e is assumed to be
Gaussian white noise with standard deviation
. As is well
known, the restoration of
f
from its observation
is
mathematically an ill-posed problem. In general, a common
variational regularization method is used to reformulate
image restoration as a well-posed problem.
Many regularization approaches have been proposed for
image restoration, in which one of the most popular
regularizers is the total variation (TV) regularizer that
provides well ability to preserve image edges [1-4].
However, TV favors piecewise constant solutions so that it
easily gives rise to the undesired staircase effect. To reduce
the staircase effect, many second order regularizers have
been proposed for image restoration [5-10]. In [9,10], the
authors analyze some related properties about the Hessian-
matrix of image at each pixel and propose two novel
regularizers, i.e. Hessian-matrix spectral norm regularizer
and Hessian-matrix Frobenius norm regularizer, which give
better performance in the process of natural image and
biomedical image restoration. More specially, the Hessian-
matrix Frobenius norm regularization image restoration
model is formulated as following optimization problem:
¿
¾
½
¯
®
+−
³
Ω
)(
2
min
2
fJdxdyfAg
F
f
λ
, (2)
where
0≥
is the regularization parameter, and )( fJ
F
is the Hessian-matrix Frobenius norm regularizer which is
defined as
dxdyyxffJ
F
³
Ω
=
2
||),(V||)( , (3)
where
Τ
∂∂∂= ),2,(V
yyxyxx
denotes the second order
differential operator, and
2
|||| ⋅
denotes the Euclidean norm.
Moreover, as referred in [9,10], the authors have designed
two different algorithms for the above Hessian-matrix
Frobenius norm regularization image restoration model (2),
which are a projected gradient algorithm (PGA) and a
preconditioned conjugate gradient algorithm that are both
derived following the majorization-minimization (MM)
framework, see [9,10] for more details. Especially, as
reported in [10], PGA exhibits state-of-the-art convergence
rate, where the convergence rate of PGA fully depends on
solving the resulting denoising sub-problem that can be
solved by the proximal-point method [11,12] in the MM
framework. To further improve the convergence rate, it
motivates us to investigate faster and more efficient
algorithms outside MM framework.
In this paper, we employ variable splitting and augmented
Lagrangian method to give an equivalent constrained
formulation of model (2), and then an efficient alternating
direction minimization algorithm is presented to solve the
resulting constrained optimization problem outside the MM
framework.
The rest of this paper is organized as follows. In section 2,
2014 Sixth International Conference on Intelligent Human-Machine Systems and Cybernetics
978-1-4799-4955-7/14 $31.00 © 2014 IEEE
DOI 10.1109/IHMSC.2014.104
3