Yuan et al. / Front Inform Technol Electron Eng 2015 16(12):1069-1087
1072
et al., 2006a; 2006b; Donoho, 2006). Under certain
conditions, this l
0
-norm problem can be approxi-
mately replaced by the convex relaxation, l
1
-norm
optimization, which simplifies the solution process of
Eq. (1) and promotes sparsity. Besides, in the process
of k-space data acquisition, random noise is
unavoidable. The minimization problem in Eq. (1)
turns into
1u2
min || || s.t. || || ,
x
Ψ Fx yx (2)
where ε is a parameter which depends upon the added
noise variance. By merging the constraint term into
objective function (2), the formula turns out to be
2
u2 1
1
min || || || || ,
2
x
xFx y Ψ
(3)
where the Lagrangian multiplier
λ>0 controls the
tradeoff between solution sparsity and data fidelity. In
Eq. (3), the error term is used to constrain the con-
sistency of the reconstructed image with
k-space data,
and the sparse constraint term is used to guarantee the
sparsity in the transform domain.
As pointed out by Afonso
et al. (2011), parameter
in Eq. (2) has a straightforward meaning, which is
proportional to the noise standard deviation, and is
much easier to set than parameter
λ in Eq. (3). Con-
sequently, in this work, we focus directly on the con-
strained problems (1) and (2) by a fast algorithm.
2.2 Uniform discrete curvelet transform
UDCT (Nguyen and Chauris, 2010) is a novel
mathematical and computational tool for multi-
resolution data representation and an innovative im-
plementation of the discrete curvelet transform,
which uses the ideas of fast Fourier transform (FFT)
based discrete curvelet transform and filter-bank
based contourlet transform. The discrete curvelet
functions are defined by a parameterized family of
smooth windowed functions that satisfy two condi-
tions: they are 2π periodic and their squares form a
partition of unity, and the centers of the curvelet
functions at each resolution are positioned on a uni-
form lattice. UDCT is implemented by the FFT algo-
rithm but designed as a multi-resolution filter-bank
with the advantages of the two methods.
Compared with other directional, discrete, and
nonadaptive transforms, UDCT provides a flexible
instead of fixed number of directions at each level to
accurately capture various directional geometrical
structures of the image. UDCT has several advantages
over existing transforms in practical applications,
such as lower redundancy ratio, hierarchical data
structure, and ease of implementation. These make
UDCT very practical in many applications. Further-
more, its shift-invariance in the energy sense is sig-
nificant in image analysis and representation. Fig. 1
illustrates some effective atoms learned from different
UDCT coefficient sub-bands of some scales after
UDCT operating on the image. For one sub-dictionary,
all coefficients are set to zeros except one in terms of
this sub-dictionary, to visualize a single effective
atom of it. Passing the result of multiplying such a
coefficient set by the learned multi-scale dictionary,
through the uniform discrete curvelet synthesis oper-
ation, exhibits a visualization of a single ‘effective’
atom in the image domain, demonstrating that the
atoms are localized and from different scales, possess
clear directionality, and are adapted to the training
data.
Fig. 1 Visualization of some effective atoms from different
levels/bands trained on an undersampling T2-weighted
image of the brain using a four-level UDCT
A separate sub-dictionary was trained for each band. Atoms
came from the approximation band (a), the first direction of
the second level (b), the first direction of the third level (c),
and six directions of the fourth level (d)–(i)
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)