MPE Mathematical Problems in Engineering 1563-5147 1024-123X Hindawi Publishing Corporation 790547 10.1155/2014/790547 790547 Research Article TV+TV2 Regularization with Nonconvex Sparseness-Inducing Penalty for Image Restoration Lu Chengwu Huang Hua Yang Suh-Yuh Key Laboratory of Data Analysis and Image Processing Chongqing University of Arts and Sciences Chongqing China international.cqu.edu.cn 2014 432014 2014 24 09 2013 02 01 2014 16 01 2014 4 3 2014 2014 Copyright © 2014 Chengwu Lu and Hua Huang. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

In order to restore the high quality image, we propose a compound regularization method which combines a new higher-order extension of total variation (TV+TV2) and a nonconvex sparseness-inducing penalty. Considering the presence of varying directional features in images, we employ the shearlet transform to preserve the abundant geometrical information of the image. The nonconvex sparseness-inducing penalty approach increases robustness to noise and image nonsparsity. In what follows, we present the numerical solution of the proposed model by employing the split Bregman iteration and a novel p-shrinkage operator. And finally, we perform numerical experiments for image denoising, image deblurring, and image reconstructing from incomplete spectral samples. The experimental results demonstrate the efficiency of the proposed restoration method for preserving the structure details and the sharp edges of image.

1. Introduction

Image restoration is one of the most classical problems in image processing. During the formation, transmission, or recording process, images may be degraded by noise and blur. Thus, an important problem in image processing is the restoration of the original true image from an observed image. Mathematically, it amounts to estimate an image uN from a measurement (1)f=𝒜u+w obtained through a noninvertible linear degradation operator 𝒜:NM (i.e., an M×N matrix) and contaminated by an additive noise wM. Typical degradation operators include identity as in image denoising, incomplete linear projector as in compressive sensing, convolution operator as in image deconvolution problems, 0-1 mask as in image inpainting, and so on.

The recovery of u from the observed image fM can be considered as an inverse problem which is ill-posed in general. The information provided by f is not sufficient to ensure existence, uniqueness, and stability of a solution u. It is therefore necessary to regularize the problem by adding a prior constraint on the solution. One typical regularization approach takes the following form: (2)minuN12𝒜u-f22+λ(u), where the first term is the fidelity function (or fitting function) that forces closeness to data according to (1), :N- is the so-called regularization function (or penalty function) that embodies the priors about the unknown image u with certain characteristics (such as piecewise smoothness or sparseness), and λ>0 is a parameter that controls the trade-off between these two functions. The intuitive meaning of the linear inverse problem (2) is clear: its minimizers reach a compromise between lack of fitness to the observed image (as measured by the fidelity function) and a degree of “undesirability” (as quantified by (·)).

The choice of the regularization function (·) significantly affects the quality of the restored image. The majority of the existing regularization functions favor simpler images with piecewise constant intensities. Details and fine features in images are not necessarily more complicated. The existing state-of-the-art image restoration models usually are designed for single convex regularization function such as total variation (TV) seminorm or sparseness-inducing penalty. The success and widespread use of the TV seminorm in the past two decades can be mainly attributed to its ability to produce well-preserved and sharp edges. Moreover, its convexity permits the design of efficient algorithms. Although the TV seminorm is extremely popular in a variety of applications, it also gives rise to some undesired effects. In particular, it leads to the well-known staircase effect when applied to images that are not necessarily piecewise constant . To attenuate the staircase effect, there is a growing interest in the literature for replacing TV seminorm by combining a second-order regularization function with the TV seminorm  or employing a 1 norm of some higher-order differential operator in a standalone way . The motivations behind these attempts are to restore potentially a wider range of images, which comprise more than merely piecewise constant regions. More recently, Papafitsoros and Schönlieb constitute a straightforward higher-order extension of the TV seminorm by adding a nonsmooth second-order regularization function; it is referred to as TV+TV2 regularization . TV+TV2 regularization can compete with total generalized variation (TGV) , infimal convolution , and Euler’s elastic , three other state-of-the-art higher-order regularizations. But most of all, TV+TV2 is simple and efficiently numerically solvable.

On the other hand, problems of taking (·) as sparseness-inducing penalty have become familiar over the past three decades, particularly in statistical and signal processing contexts. One typical sparseness-inducing penalty problem takes the following form: (3)minuN12𝒜u-f22+λΨu1, where ·1 denotes the 1 norm and Ψ is a sparsity operator, such as a wavelet transform. When Ψ is a discrete gradient operator, the second term leads to TV regularization. In the 1990s, seminal work on the use of 1 sparseness-inducing penalty appeared in the statistics literature, such as the famous basis pursuit denoising (BPDN) criterion  and the least absolute shrinkage and selection operator (LASSO, ). For brief historical accounts on the use of the 1 penalty in statistics and signal processing, see . Another intriguing new application for the 1 sparseness-inducing penalty is compressive sensing (CS)  when 𝒜 is the measurement or sampling operator in the optimization problem (3). In addition, it is worth mentioning the numerical results in [19, 20] and theoretical works . These results have justified that the robustness of restored images to noise and the nonsparsity can be increased by replacing 1 norm in (3) with the nonconvex nonsmooth p(0<p<1) quasinorm. Moreover, recent researches show that p(0<p<1) quasinorm offers much richer possibilities to restore high quality images with neat edges.

To restore high quality image, in the regularization approach (2), one may be desired to encourage the solution to exhibit characteristics that are piecewise smooth and sparse simultaneously. The resulting optimization problem cannot be dealt with by single regularization. It is most naturally expressed by the combination of more than one regularization function, such as compounding total variation and sparseness-inducing penalty .

In this paper, we consider a new compound regularization, that is, linear combination of TV+TV2 regularization and sparseness-inducing p(0<p<1) quasinorm. Considering the presence of varying directional features in image, we employ the shearlet transform to preserve the abundant geometrical information of the image. As mentioned above, TV+TV2 regularization can effectively attenuate the staircase effect of TV regularization. The nonconvex sparseness-inducing penalty approach increases robustness to noise and image nonsparsity.

2. Preliminary

In this section, we will briefly review TV+TV2 regularization and shearlet transform to make this paper self-contained.

2.1. TV+TV<sup>2</sup> Regularization

In image restoration model (2), the most frequently chosen regularization term is TV seminorm (4)(u)=TV(u)=Ω|u|dxdy, where |u|=:|xu|2+|yu|2. It is well-known that TV seminorm is tuned towards the preservation of edges and performs very well if the reconstructed image is piecewise constant. However, as an image does not only consist of flat regions and jumps but also possess slanted regions, that is, piecewise linear parts, such regularization procedure will cause undesired staircase effect. One way to reduce this staircase effect is in fact to “round off" TV seminorm by using Huber regularization . However, such procedure can only eliminate the staircase effect in areas with small gradient. Another way of improving TV seminorm is the introduction of higher-order derivatives in the regularization function. Especially in recent years, higher-order versions of nonsmooth image enhancing methods have been considered.

To reduce the staircase effect of TV regularization, Chambolle and Lions  understood the image u as an addition of two image layers, that is, u=u1+u2 (u1 is the piecewise constant part of u and u2 is the piecewise smooth part), and proposed the inf-convolution regularization function: (5)(u)=Ω(α|u1|+β|2u2|)dxdy, where |u1|=:|xu1|2+|yu1|2 and |2u2|=:|xxu2|2+|yyu2|2+2|xyu2|2 is the distributional Hessian of u2. We denote TV(u1)=Ω|u1|dxdy and TV2(u2)=Ω|2u2|dxdy. The former is TV regularization and the latter is TV2 regularization, which is total variation of the gradient. In [3, 7, 2628] the authors also consider using TV2 regularization.

Following the idea of the inf-convolution regularization, Chan et al. proposed the following regularization : (6)(u)=Ω(α|u1|+β|Δu2|2)dxdy,p=1,2, where the Hessian 2 used in the inf-convolution regularization is replaced by the Laplacian Δ.

Another attempt to combine first- and second-order regularization originates from Chan et al. , who consider TV regularization together with weighted versions of the Laplacian. More precisely, they proposed the following regularization: (7)(u)=Ω(α|u|+βg(|u|)|Δu|2)dxdy, where g(·) must be a function with certain growth conditions at infinity in order to allow jumps.

It is particularly worth mentioning that, in , Bredies et al. proposed the following total generalized variation (TGV), which is recognized as the currently best higher-order regularization function: (8)(u)=αu-v1+β·v1, where the variable v=(v1,v2) varies in the space of symmetric tensors of order 2 and (9)·v1:=Ω(xv1)2+(yv1)2+(xv2)2+(yv2)2dxdy.

The TGV regularization generalizes the inf-convolution regularization in the sense that it coincides with the inf-convolution regularization when v only varies in the range of . The TGV regularization improves the results for image denoising over TV regularization and inf-convolution regularization. The interested reader can consult  for details.

More recently, in , the authors choose the following combined first- and second-order regularization function: (10)(u)=Ω(αϕ1(u)+βϕ1(2u))dxdy, where ϕ1:2+ and ϕ2:4+ are all convex function with at most linear growth at infinity. When one considers ϕ1(x)=|x| and ϕ2(x)=|x|, the regularization procedure essentially is a combination of the total variation and the total variation of the gradient; it is referred to as TV+TV2 regularization. The idea of TV+TV2 regularization is to regularize with a fairly large weight α in the first-order term—preserving the jumps as good as possible—and use a small weight β for the second-order term such that staircase artifacts created by the first-order regularization are eliminated without introducing any serious blur in the reconstructed image.

In , the authors show that for image denoising, deblurring, and inpainting TV+TV2 regularization offers solutions whose quality is not far off from the ones produced by TGV regularization. Moreover, the computational effort needed for its numerical solution is not much more than the one needed for solving TV regularization . For comparison the numerical solution for TGV regularization is in general about ten times slower than TV+TV2 regularization.

In addition, it is well-known that TV regularization is not optimal for the restoration of textured regions. In recent years directional representation systems were proposed to cope with curved singularities in images. In particular, curvelets and shearlets provide an optimally sparse approximation in the class of piecewise smooth functions with C2 singularity boundaries. In the next subsection, we will briefly review the discrete shearlet transform that is suited as sparsity operator for the restoration of curved structures.

2.2. Shearlet Transform

The classical wavelet transform is unable to provide additional information about the geometry of the set of singularities of image. This is due to the fact that this transform is isotropic. As a result, it has a very limited ability to resolve edges and other distributed discontinuities which usually are very important for image. Shearlet transform is a directional representation system that provides more geometrical information for multidimensional functions such as images .

Let ψL2(2), Aa=[a00a]  (a+) is parabolic scaling matrix, and Ss=[1s01] (s) is shear matrix. The shearlet system {ψast,a+,s,t2} is generated by applying the operations of dilation, shear transformation, and translation on ψ: (11)ψast(x):=a-3/4ψ(Aa-1Ss-1(x-t)).

The continuous shearlet transform of function fL2(2) is defined as (12)Sψ(f)(a,s,t)=f,ψast.

The shearlet transform is invertible if the function ψ satisfies the admissibility property (13)2|ψ^(ω1,ω2)|2|ω1|2dω1dω2, where ψ^ is Fourier transform of ψ.

To obtain a discrete shearlet transform, we consider digital image as function sampled on the grid (1/N):(1/N){(m1,m2):mi=0,,N-1,i=1,2}, where we assume quadratic image for simplicity. Further, let j0:=(1/2)log2N be the number of considered scales. We discretize the scaling, shear, and translation parameters as (14)aj:=2-2j=14j,j=0,,j0-1,sj,k:=k2-j,-2jk2j,tm:=mN,m. With these notations the shearlet becomes ψj,k,m:=ψaj,sk,tm(x)=ψ(Aaj-1Ssj,k-1(x-tm)).

We select shearlet in this paper due to its directional sensitivity, availability of efficient implementation. Shearlet transform is by far the most optimal sparsifying transform for piecewise smooth image containing rich geometric information such as edges, corners, and spikes. It combines the power of multiscale methods with the ability of extracting geometry of image.

3. Our Model and Algorithm 3.1. Unconstrained Formulation for Our Model

In this subsection, we present our new model as follows that employs TV+TV2 regularization with a nonconvex sparseness-inducing penalty based on shearlet transform: (15)minuN12𝒜u-f22+αΩ|u|dx+βΩ|2u|dx+γSψ(u)pp, where 0<p<1. This is a nonconvex optimization problem. One approach to handle this kind of nonconvex optimization problem is the iteratively reweighted least-squares (IRLS) approach which was previously taken by Rao and Kreutz-Delgado , but with unimpressive results due to getting trapped in local minima. The alternative herein proposed is to use an operator splitting method and the Bregman iteration method, as well as a novel p-shrinkage operator to solve (15).

To facilitate the discussion of our algorithm, without the loss of generality, we represent a gray image as n×n matrix, and the Euclidean space n×n is denoted as V, so uV, fV. The linear operator 𝒜 is a mapping 𝒜:VV. Using the above notations, the model (15) can be expressed as follows: (16)minuV12𝒜u-f22+αu1+β2u1+γSd(u)pp, where Sd(u)=(S1(u),S2(u),,SJ(u)) is the discrete shearlet transform of u and Sj(u)(j=1,2,,J) denotes its jth subband. u=(Dxu,Dyu) and 2u=(Dxxu,Dyyu,Dxyu,Dyxu) are the discrete differential operators of first- and second-order, respectively. By assuming periodic boundary conditions for u, we define Dxu and Dyu as a forward difference operator, respectively, as follows: (17)(Dxu)i,j=(Dx+u)i,j={ui,j+1-ui,j1jn-1ui,1-ui,nj=n,(Dyu)i,j=(Dy+u)i,j={ui+1,j-ui,j1in-1u1,j-un,ji=n. We also need the discrete divergence operator div:V2V that has the adjointness property (18)-divq·u=q·u,uV,qV2. This property forms the discrete analogue of integration by parts. For q=(q1,q2)V2 we define (19)(divq)i,j=(Dx-q1)i,j+(Dy-q2)i,j, where Dx- and Dy- are the following backward difference operators: (20)(Dx-u)i,j={ui,j-ui,j-12jnui,1-ui,nj=1,(Dy-u)i,j={ui,j-ui-1,j2inu1,j-un,ji=1.

Analogously we define the second-order discrete differential operators Dxxu, Dyyu, and Dxyu. These are just the corresponding compositions of the first-order operators: (21)Dxxu=Dxx+u={ui,n-2ui,j+ui,j+1j=1ui,j-1-2ui,j+ui,j+11<j<nui,j-1-2ui,j+ui,1j=n,(22)Dxxu=Dyy+u={un,j-2ui,j+ui+1,ji=1ui-1,j-2ui,j+ui+1,j1<i<nui-1,j-2ui,j+u1,ji=n,(23)Dxyu=Dxy+u={ui,j-ui+1,j-ui,j+1+ui+1,j+11i<n,1j<1ui,j-u1,j-ui,j-1-2ui,j+u1,j+1i=n,1j<nui,j-ui+1,j-ui,1+ui+1,11i<n,j=nui,j-u1,j-ui,1+u1,1i=n,j=n.

One can easily check that Dxy+=Dx+(Dy+)=Dy+(Dx+)=Dyx+. As before, we also need the discrete second-order divergence operator div2:V4V with the adjointness property (24)div2q·u=q·2u,uV,qV4.

For a q=(q11,q22,q12,q21)V4 we define (25)(div2q)(i,j)=Dxx-q11(i,j)+Dyy-q22(i,j)+Dxy-q12(i,j)+Dyx-q21(i,j), where Dxx-=Dxx+, Dyy-=Dyy+, and Dxy-=Dyx- with (26)Dxy-u={ui,j-ui,n-un,j+un,ni=1,j=1ui,j-ui,j-1-un,j+un,j-1i=1,1<jnui,j-ui-1,j-ui,n+ui-1,n1<in,j=1ui,j-ui,j-1-ui-1,j+ui-1,j-11<in,1<jn.

Remark 1.

By choosing periodic boundary conditions, the action of each of the discrete differential operators can be regarded as a circular convolution of u and allows the use of fast Fourier transform, which our algorithm is relying on.

3.2. Constrained Formulation for Our Model

In the following we use what is frequently called the split Bregman iteration to solve (16); see [33, 34]. To do so, we firstly formulate the unconstraint minimization problem (16) as a constraint problem and then apply the Bregman iteration to solve it. We introduce auxiliary variables v=(v1,v2)V2, w=(w1,w2,w3,w4)V4, and c=(c1,c2,,cJ)VJ such that (16) is equivalent to the following constrained minimization problem: (27)minuV,vV2,wV4,cVJ12𝒜u-f22+αv1+βw1+γcpps.t.v=u,w=2u,c=Sd(u).

Here, we define (28)𝒜u-f2=(i=1nj=1n(𝒜u(i,j)-f(i,j))2)1/2v1=i=1nj=1n(v1(i,j)2+v2(i,j)2)1/2w1=i=1nj=1n(w11(i,j)2+w22(i,j)2+w12(i,j)2+w21(i,j)2)1/2cp=(i=1nj=1nl=1Jcl(i,j)p)1/p.

Note that the problems (16) and (27) are equivalent. The Bregman iteration that corresponds to (27) is the following: (29)[uk+1,vk+1,wk+1,ck+1]=arg minu,v,w,c12𝒜u-f22+αv1+βw1+γcpp+λ12b1k+u-v22+λ22b2k+2u-w22+λ32b3k+Sd(u)-c22,(30)b1k+1=b1k+uk+1-vk+1,(31)b2k+1=b2k+2uk+1-wk+1,(32)b3k+1=b3k+Sd(uk+1)-ck+1, where b1k=(b1,1k,b1,2k)V2, b2k=(b2,1,1k,b2,2,2k,b2,1,2k,b2,2,1k)V4, b3k=(b3,1k,b3,2k,,b3,Jk)VJ, and λ1, λ2, λ3 are positive constants.

It is difficult to solve the minimization problem (29) directly; thus we split it into four separate subproblems for u, v, w, and c, each of which can be solved fast as follows.

Subproblem  1: (33) u k + 1 = arg min u 1 2 𝒜 u - f 2 2 + λ 1 2 b 1 k + u - v k 2 2 + λ 2 2 b 2 k + 2 u - w k 2 2 + λ 3 2 b 3 k + S d ( u ) - c k 2 2 .

Subproblem  2: (34) v k + 1 = arg min v α v 1 + λ 1 2 b 1 k + u k + 1 - v 2 2 .

Subproblem  3: (35) w k + 1 = arg min w β w 1 + λ 2 2 b 2 k + 2 u k + 1 - w 2 2 .

Subproblem  4:   (36) c k + 1 = arg min c γ c p p + λ 3 2 b 3 k + 𝒮 d ( u k + 1 ) - c 2 2 .

We now describe how we solve each of the four subproblems (33)–(36).

3.3. To Solve Subproblem 1

Subproblem 1 can be solved through its optimality condition. The optimality condition reads as follows: (37)u(12𝒜u-f22+λ12b1k+u-vk22+λ22b2k+2u-wk22+λ32b3k+Sd(u)-ck22)=0.

The jth subband of shearlet transform of u is able to be efficiently implemented as component-wise multiplication with a mask matrix denoted by Hj in frequency domain: (38)Sj(u)=-1Hju=MHju with MHl=-1Hj. Here and -1 are the 2D discrete Fourier transform and inverse Fourier transform, respectively.

Now by solving the above optimality condition, we get (39)𝒜*𝒜u-λ1(Dx-Dx++Dy-Dy+)u+λ2(Dxx-Dxx++Dyy-Dyy++2Dxy-Dxy+)u+λ3l=1JMHl*MHlu=𝒜*f+λ1(Dx-Dx+(b1,1k-v1k)+Dy-Dy+(b1,2k-v2k))-λ2(Dxx-Dxx+(b2,1,1k-w1,1k)+Dyy-Dyy+(b2,2,2k-w2,2k)+2Dxy-Dxy+(b2,1,2k-w1,2k))-λ3j=1JMHj*(b3,jk-cjk). Taking into account that the discrete differential operators act like circular convolutions, we can use the 2D discrete Fourier transform to solve system (39) which has a closed-form solution (40)FD·(uk+1)=(Rk), where “·” means pointwise multiplication of matrices. Consider (41)FD(𝒜*𝒜)-λ1((Dx-(Dx+))+(Dy-(Dy+)))+λ2((Dxx-(Dxx+))+(Dyy-(Dyy+))+(Dxy-(Dxy+)))+λ3((j=1JMHj*MHj))=(𝒜*𝒜)-λ1((Dx-(Dx+))+(Dy-(Dy+)))+λ2((Dxx-(Dxx+))+(Dyy-(Dyy+))+(Dxy-(Dxy+)))+λ3j=1J|Hj|2 and Rk is the right hand side of (39). The term FD needs to be calculated only once at the beginning of the algorithm. Thus, we have (42)uk+1=-1((Rk)FD).

3.4. To Solve Subproblems 2 and 3

Subproblems 2 and 3 are similar and the solutions can be explicitly expressed in the form of shrinkage (43)vk+1=Shink(b1k+uk+1,αλ1),(44)wk+1=Shink(b2k+2uk+1,βλ2), where Shink(t,σ)=sgn(t)max(|t|-σ,0).

3.5. To Solve Subproblem 4

Now we present the p-shrinkage operator to solve subproblem 4. To avoid nondifferentiability caused by the p penalty term in subproblem 4, we consider its smooth approximation problem (45)ck+1=arg mincγφτ(c)+λ32b3k+𝒮d(uk+1)-c22, where φτ(·) is defined by φτ(t)={κ|t|2if|t|τ|t|p/p-δif|t|>τ, for τ>0 and any real p. Here the parameters κ and δ are chosen to make φτ(·) a C1 function, and τ will be chosen shortly. For convenience we adopt the unusual convention that, when p=0, |t|p/p will mean log(|t|).

Next, we show that there is a function ψτ satisfying (46)φτ(t)=mins{ψτ(s)+η2s-t22}. Let (47)θτ(t)=|t|22-1ηφτ(t),ϑτ(s)=|s|22+1ηψτ(s);φτ(t) is a smooth approximation of tpp; it still is nonconvex, but by taking τη1/(p-2), above the auxiliary function θτ(t) is convex. Simple manipulation shows that, for φτ and ψτ to satisfy (46), it is necessary and sufficient to have θτ=ϑτ*, where ϑτ* is the conjugate (or Legendre-Fenchel transform) of ϑτ defined as (48)ϑτ*(s)=supt{s,t-ϑτ(t)} which shows how to construct ψτ from φτ through computing ϑτ from θτ. Simple manipulations then show that (46) is satisfied by (49)ψτ(s)=η(ϑτ(s)-|s|22). Moreover, the minimizer of (46) will be given by s*=θτ(t), which can be shown to be given by what we call the p-shrinkage operator as follows: (50)Sτp(t)=sgn(t)max{|t|-τ|t|p-1,0}. This shrinkage operator will not be necessary to compute ψτ(s) explicitly. We now incorporate the above into our algorithm. By using the half-quadratic technique originally introduced by Geman and Yang in , the approximation problem (45) becomes an equivalent augmented problem as follows: (51)ck+1=arg minc,sψτ(s)+η2s-c22+λ32γb3k+𝒮d(uk+1)-c22, where s=(s1,s2,,sJ)VJ.

By letting η, s is forced to approach c, making the solution to (51) approach that of (45). There is no interaction between c and s in (51), so (51) can become the following two subproblems which are alternately computed: (52)sk+1=arg minsψτ(s)+η2s-ck22,(53)ck+1=arg mincη2sk+1-c22+λ32γb3k+𝒮d(uk+1)-c22.

Ma et al.  use a continuation approach to solve (52), starting with small η values and then increasing them geometrically, using the solution at each stage to initialize the next stage. Instead, we adopt the Bregman iteration approach of Goldstein and Osher .

For fixed ck, the resulting algorithm for (52) is as follows: (54)sk+1=S1/ηp(ck+b4k),b4k+1=b4k+ck-sk+1, where b4k=(b4,1k,b4,2k,,b4,Jk)VJ.

For fixed sk=1, (53) has a closed-form solution: (55)ck+1=ηγηγ+λ3sk+1+λ3ηγ+λ3(b3k+Sj(uk+1)).

The resulting algorithm is summarized in Algorithm 1.

<bold>Algorithm 1: </bold>TV+TV<sup>2</sup>+<inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M209"><mml:mrow><mml:msub><mml:mrow><mml:mi>ℓ</mml:mi></mml:mrow><mml:mrow><mml:mi>p</mml:mi></mml:mrow></mml:msub></mml:mrow></mml:math></inline-formula> sparseness-inducing penalty.

Initialisation:

(1)  uk+1 is given by (42) for solving Subproblem 1;

(2) vk+1 is given by (43) for solving Subproblem 2;

(3) wk+1 is given by (44) for solving Subproblem 3;

(4) ck+1 is given by (54) and (55) for solving Subproblem 3;

(5) Update on b1: b1k+1=b1k+uk+1-vk+1;

(6) Update on b2: b2k+1=b2k+2uk+1-wk+1;

(7) Update on b3: b3k+1=b3k+Sd(uk+1)-ck+1.

4. Numerical Experiments

In this section, we will perform some image restoration experiments using our proposed TV+TV2+p sparseness-inducing penalty model (16). The image restoration scenarios will be denoising, deblurring, and reconstruction from incomplete data. We will use Algorithm 1 to solve (16). We compare our method with the one proposed in  (TV method) and the one proposed in  (TV+TV2 method). All computations of the present paper are carried out in Matlab 7.0. The results are obtained by running the Matlab codes on an Intel(R) Core (TM) i3 CPU (3.20 GHz, 3.19 GHz) computer with RAM of 1.92 GB.

The quality of the restoration results with different methods is compared quantitatively by using the peak signal-to-noise ratio (PSNR), the mean-squared error (MSE), and the structural similarity index (SSIM). They are defined as follows: (56)PSNR=10·log10(255×255MSE),MSE=1mnu^-u22, where u and u^ are the expected image of size m×n and the restored one, respectively, and (57)SSIM=(2μ1μ2+C1)(2σ12+C2)(2μ12μ22+C1)(σ12+σ22+C2), where μ1 and μ2 are averages of u and u^, respectively, σ1 and σ2 are the variance of u and u^, respectively, and σ12 is the covariance of u and u^. The positive constants C1 and C2 can be thought of as stabilizing constants for near-zero denominator values. The SSIM is a well-known quality metric used to measure the similarity between two images. The SSIM metric was developed by Wang et al.  and is based on three specific statistical measures that are much closer to how the human eye perceives differences between images. The SSIM index also assesses the conservation of the structural information of the reconstructed image. Note that the perfect reconstruction would have SSIM value equal to 1. In the following experiments, we will use SSIM map to reveal areas of high or low similarity between two images, the whiter SSIM map, and the closer between the two images.

4.1. Denoising Experiments

In our denoising implementation, the linear operator 𝒜 equals the identity. We perform experiments to images that are corrupted a white Gaussian noise with standard deviation σ. Note that, for α0, β=0, and γ=0, our model (16) corresponds to the classical Rudin-Osher-Fatemi denoising model , and, for α0, β0, and γ=0, it corresponds to the TV+TV2 restoration model . In the following we will examine whether the introduction of the higher-order term (β0) and the nonconvex sparseness-inducing penalty term (γ0) in the denoising procedure produces results of higher quality.

We test our method on the cropped Lena image of size 256×256. The image is contaminated with Gaussian noise of variance 0.05. For comparison, we report the results of our proposed method (taking parameters α=0.02, β=0.01, γ=0.05, λ1=0.2, λ2=0.1, λ3=0.0001, and μ=1000), the TV regularization method (taking parameters α=0.02, β=0, γ=0, λ1=0.2, λ2=0, and λ3=0), and the TV+TV2 regularization method (taking parameters α=0.02, β=0.01, γ=0, λ1=0.2, λ2=0.1, and λ3=0). In general, we have found that λ1=10α, λ2=10β, and p0.5 usually result in good convergence. Iterations are terminated when the condition uk+1-uk2/uk+1210-3 is met. The denoising results by all three methods are shown in Figures 1 and 2. For better visualization, we include the middle row slices and the patch 3D profile. Note that the nonconvex parameter is p=0.5 for our proposed method. We also have found that the convergence speed will become slow if p<0.5, and the results are roughly the same if p0.5. The following deblurring and reconstructing experiments are also this kind of circumstance. Compared with the other methods, the proposed method yields better visual results since it is able to preserve edges and fine features and provide more “natural-looking” images.

Comparison of Gaussian noise removal. First column from up to down: original image, noisy image (PSNR = 58.1704, MSE = 0.0998, and SSIM = 0.4210), results of the TV method (PSNR = 63.3717, MSE = 0.0521, and SSIM = 0.8224), results of the TV+TV2 method (PSNR = 63.9644, MSE = 0.0298, and SSIM = 0.8512), and results of our proposed TV+TV2+p method (PSNR = 64.1546, MSE = 0.0161, and SSIM = 0.8907). Second column: middle row slices on the blue line. Third column: 3D close-up intensity profile in the red box. Fourth column: SSIM map.

Denoising performance comparisons on the Lena image with different method.

Iterations versus PSNR (dB)

Iterations versus SSIM

4.2. Deblurring Experiments

In our deblurring implementation, 𝒜 denotes a circular convolution with a discrete approximation of a Gaussian kernel of size 11×11. As an additional degradation factor, the initial image is corrupted by additive Gaussian noise of variance 0.001. The quality of the resulting images is measured in terms of an increase in the signal-to-noise ratio (ISNR) measured in dB. The ISNR is defined as (58)ISNR=10log10(MSE1MSE2), where MSE1 and MSE2 are the mean-squared error between the degraded and the original image and the restored and the original image, respectively. The test image takes the Goldhill image of size 256×256. In this experiment, we take α=0.00001, β=0.00002, γ=0.001, λ1=0.0001, λ2=0.0002, λ3=0.0001, and μ=5000 for our method, α=0.00001, β=0, γ=0, λ1=0.0001, λ2=0, and λ3=0 for the TV method, and α=0.00001, β=0.00002, γ=0, λ1=0.0001, λ2=0.0002, and λ3=0 for the TV+TV2 model. As the previous experiments, in these deblurring experiments, we still take p=0.5 for our proposed method.

For visual appreciation, the corresponding restoration results are reported in Figures 3 and 4. The results show the consistent advantage of the proposed method over all the others. The TV solution is characterized by sharper edges, as expected, but also results in staircase effects appearing in the smoother regions of the image. The TV+TV2 solution produces the smoother results and loses some details of the image. The result of our method achieves better visual effect and quantitative index than other methods.

Comparison of the blurry and noisy image restoration. First column from up to down: original image, degraded image (PSNR = 59.0771, ISNR = 0.0, MSE = 0.0804, and SSIM = 0.4112), results of the TV method (PSNR = 62.2991, MSE = 0.0401, ISNR = 3.2220, and SSIM = 0.6907), results of the TV+TV2 method (PSNR = 62.8713, MSE = 0.0344, ISNR = 3.7942, and SSIM = 0.8149), and results of our proposed TV+TV2+p method (PSNR = 63.1684, MSE = 0.0286, ISNR = 4.5042, and SSIM = 0.8547). Second column: close-up image in the red box. Third column: 3D close-up intensity profile in the red box. Fourth column: SSIM map.

Deblurring performance comparisons on the Goldhill image with different method.

Iterations versus ISNR (dB)

Iterations versus SSIM

4.3. Reconstructing Experiments from Incomplete Data

Now we show the performance of the proposed method on the incomplete spectral data. Here, the linear operator 𝒜 denotes the incomplete measurements on the Fourier transform domain of image. To simulate the incomplete measurements, we start with a clean image, normalize its intensity values to [0,1], and then apply discrete Fourier transform to the clean image. The incomplete measurements are obtained by keeping the samples in some radial lines passing through the center in Fourier domain. In our reconstructing experiments, we use Shepp-Logan phantom image and Brain image as the test images. We consider 22 radial lines k-space sampling (9.36% DCT data). The reconstructing parameters are taken as α=0.3, β=0.01, γ=0.01, and λ3=0.005, for our method, α=0.3, β=0, γ=0, and λ3=0 for the TV method, and α=0.3, β=0.01, γ=0, λ1=0.03, λ2=0.001, and λ3=0 for the TV+TV2 method.

The corresponding reconstruction results are reported in Figures 5, 6, and 7. Because the phantom has a very sparse gradient, all methods obtain good visual effect in the reconstruction of Shepp-Logan phantom image from 9.36% DCT data. Although our proposed method achieves a better quantitative index than the TV and the TV+TV2, the results look very similar and it is difficult to observe any significant visual differences. However, the proposed method demonstrates the distinct advantage in the reconstruction of the Brain image.

Comparison of Shepp-Logan phantom image reconstruction from 9.36% DCT data. First column from up to down: original image, back projection image (PSNR = 56.9194, MSE = 0.1322, and SSIM = 0.2790), results of the TV method (PSNR = 65.1352, MSE = 0.0274, and SSIM = 0.9768), results of the TV+TV2 method (PSNR = 65.3658, MSE = 0.0189, and SSIM = 0.9820), and results of our proposed method (PSNR = 66.6062, MSE = 0.0142, and SSIM = 0.9896). Second column: close-up image in the red box. Third column: 3D close-up intensity profile in the red box. Fourth column: SSIM map.

Comparison of reconstruction of Brain image from 9.36% DCT data. First column from up to down: original image, back projection image (PSNR = 58.0251, MSE = 0.1025, and SSIM = 0.3919), results of the TV method (PSNR = 59.8861, MSE = 0.0668, and SSIM = 0.5376), results of the TV+TV2 method (PSNR = 62.5794, MSE = 0.0395, and SSIM = 0.7519), and results of our proposed TV+TV2+p method (PSNR = 64.4744, MSE = 0.0232, and SSIM = 0.8713). Second column: close-up image in the red box. Third column: 3D close-up intensity profile in the red box. Fourth column: SSIM map.

Reconstruction performance (Iterations versus MSE) comparisons on two test images.

Shepp-Logan phantom image

Brain image

5. Conclusion

We present a TV+TV2+p sparseness-inducing penalty image restoration model. The TV+TV2 regularization combines convex functions of the total variation and the total variation of the first derivatives. It leads to a significant reduction of the staircase effects commonly seen in traditional TV based image processing algorithms. The introduction of the nonconvex p sparseness-inducing penalty based on the shearlet transform increases robustness to image nonsparsity. Our algorithm makes use of previous operator splitting methods and the Bregman iteration method, as well as a novel p-shrinkage operator. Numerical results show the consistent excellence of our proposed method in preserving diverse details during the restoration.

Future researches aim at investigating the theoretical rules on selecting the regularization parameters and the nonconvex parameter p(0<p<1) provided by our model.

Conflict of Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Acknowledgments

The authors would like to thank the National Natural Science Foundation of China (no. 61271452), the Key Project of Chinese Ministry of Education (no. 2011155), and the Chongqing Natural Science Foundation (no. 2011jjA40029) for supporting this work. They thank Rick Chartrand and Weihong Guo for providing them with very useful suggestions that improved the presentation of the paper.

Yuan J. Schnorr C. Steidl G. Total-variation based piecewise affine regularization Scale Space and Variational Methods in Computer Vision 2009 5567 552 564 Lecture Notes in Computer Science 10.1007/978-3-642-02256-2_46 Chan T. Marquina A. Mulet P. High-order total variation-based image restoration SIAM Journal on Scientific Computing 2000 22 2 503 516 10.1137/S1064827598344169 MR1780611 Lysaker M. Tai X.-C. Iterative image restoration combining total variation minimization and a second-order functional International Journal of Computer Vision 2006 66 1 5 18 2-s2.0-29344446686 10.1007/s11263-005-3219-7 Chambolle A. Lions P.-L. Image recovery via total variation minimization and related problems Numerische Mathematik 1997 76 2 167 188 10.1007/s002110050258 MR1440119 Chan T. F. Esedoglu S. Park F. E. Image decomposition combining staircase reduction and texture extraction Journal of Visual Communication and Image Representation 2007 18 6 464 486 2-s2.0-35448969357 10.1016/j.jvcir.2006.12.004 You Y.-L. Kaveh M. Fourth-order partial differential equations for noise removal IEEE Transactions on Image Processing 2000 9 10 1723 1730 10.1109/83.869184 MR1807566 Lysaker M. Lundervold A. Tai X.-C. Noise removal using fourth-order partial differential equation with applications to medical magnetic resonance images in space and time IEEE Transactions on Image Processing 2003 12 12 1579 1590 2-s2.0-4544278119 10.1109/TIP.2003.819229 Steidl G. A note on the dual treatment of higher-order regularization functionals Computing 2006 76 1-2 135 148 10.1007/s00607-005-0129-z MR2174675 Papafitsoros K. Schönlieb C. B. A combined first and second order variational approach for image reconstruction Journal of Mathematical Imaging and Vision 2014 48 2 308 338 10.1007/s10851-013-0445-4 MR3152107 Bredies K. Kunisch K. Pock T. Total generalized variation SIAM Journal on Imaging Sciences 2010 3 3 492 526 10.1137/090769521 MR2736018 Setzer S. Steidl G. Teuber T. Infimal convolution regularizations with discrete 1-type functionals Communications in Mathematical Sciences 2011 9 3 797 827 10.4310/CMS.2011.v9.n3.a7 MR2865805 Tai X.-C. Hahn J. Chung G. J. A fast algorithm for Euler's elastica model using augmented Lagrangian method SIAM Journal on Imaging Sciences 2011 4 1 313 344 10.1137/100803730 MR2792415 Chen S. S. Donoho D. L. Saunders M. A. Atomic decomposition by basis pursuit SIAM Journal on Scientific Computing 1998 20 1 33 61 10.1137/S1064827596304010 MR1639094 Tibshirani R. Regression shrinkage and selection via the lasso Journal of the Royal Statistical Society B 1996 58 1 267 288 MR1379242 Tropp J. A. Just relax: convex programming methods for identifying sparse signals in noise IEEE Transactions on Information Theory 2006 52 3 1030 1051 10.1109/TIT.2005.864420 MR2238069 Candes E. J. Tao T. Near-optimal signal recovery from random projections: universal encoding strategies? IEEE Transactions on Information Theory 2006 52 12 5406 5425 10.1109/TIT.2006.885507 MR2300700 Candès E. J. Romberg J. Tao T. Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information IEEE Transactions on Information Theory 2006 52 2 489 509 10.1109/TIT.2005.862083 MR2236170 Donoho D. L. Compressed sensing IEEE Transactions on Information Theory 2006 52 4 1289 1306 10.1109/TIT.2006.871582 MR2241189 Chartrand R. Exact reconstruction of sparse signals via nonconvex minimization IEEE Signal Processing Letters 2007 14 10 707 710 2-s2.0-34548724437 10.1109/LSP.2007.898300 Chartrand R. Yin W. Iteratively reweighted algorithms for compressive sensing Proceedings of the 33rd IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP '08) April 2008 3869 3872 2-s2.0-51449111119 10.1109/ICASSP.2008.4518498 Chartrand R. Staneva V. Restricted isometry properties and nonconvex compressive sensing Inverse Problems 2008 24 3 1 14 035020 10.1088/0266-5611/24/3/035020 MR2421974 Saab R. Chartrand R. Yilmaz Ö. Stable sparse approximations via nonconvex optimization Proceedings of the 33rd IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP '08) April 2008 3885 3888 2-s2.0-51449102007 10.1109/ICASSP.2008.4518502 Lu C.-W. Image restoration and decomposition using non-convex non-smooth regularisation and negative Hilbert-Sobolev norm IET Image Processing 2012 6 6 706 716 10.1049/iet-ipr.2011.0345 MR3014910 Bioucas-Dias J. M. Figueiredo M. A. T. An iterative algorithm for linear inverse problems with compound regularizers Proceedings of the IEEE International Conference on Image Processing (ICIP '08) October 2008 685 688 2-s2.0-69949177588 10.1109/ICIP.2008.4711847 Pock T. Cremers D. Bischof H. Chambolle A. Global solutions of variational models with convex regularization SIAM Journal on Imaging Sciences 2010 3 4 1122 1145 10.1137/090757617 MR2763711 Scherzer O. Denoising with higher order derivatives of bounded variation and an application to parameter estimation Computing 1998 60 1 1 27 10.1007/BF02684327 MR1600531 Hinterberger W. Scherzer O. Variational methods on the space of functions of bounded Hessian for convexification and denoising Computing 2006 76 1-2 109 133 10.1007/s00607-005-0119-1 MR2174674 Lefkimmiatis S. Bourquard A. Unser M. Hessian-based norm regularization for image restoration with biomedical applications IEEE Transactions on Image Processing 2012 21 3 983 995 10.1109/TIP.2011.2168232 MR2951273 Chan T. Esedoglu S. Park F. A fourth order dual method for staircase reduction in texture extraction and image restoration problems UCLA CAM Report 2005 05-28 Rudin L. I. Osher S. Fatemi E. Nonlinear total variation based noise removal algorithms Physica D 1992 60 1–4 259 268 2-s2.0-44049111982 10.1016/0167-2789(92)90242-F Easley G. Labate D. Lim W.-Q. Sparse directional image representations using the discrete shearlet transform Applied and Computational Harmonic Analysis 2008 25 1 25 46 10.1016/j.acha.2007.09.003 MR2419703 Rao B. D. Kreutz-Delgado K. An affine scaling methodology for best basis selection IEEE Transactions on Signal Processing 1999 47 1 187 200 10.1109/78.738251 MR1729739 Goldstein T. Osher S. The split Bregman method for L1-regularized problems SIAM Journal on Imaging Sciences 2009 2 2 323 343 10.1137/080725891 MR2496060 Wu C. Tai X.-C. Augmented Lagrangian method, dual methods, and split Bregman iteration for ROF, vectorial TV, and high order models SIAM Journal on Imaging Sciences 2010 3 3 300 339 10.1137/090767558 MR2679430 Geman D. Yang C. Nonlinear image recovery with half-quadratic regularization IEEE Transactions on Image Processing 1995 4 7 932 946 2-s2.0-0029341230 10.1109/83.392335 Ma S. Yin W. Zhang Y. Chakraborty A. An efficient algorithm for compressed MR imaging using total variation and wavelets Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR '08) 2008 1 8 10.1109/CVPR.2008.4587391 Wang Z. Bovik A. C. Sheikh H. R. Simoncelli E. P. Image quality assessment: from error visibility to structural similarity IEEE Transactions on Image Processing 2004 13 4 600 612 2-s2.0-1942436689 10.1109/TIP.2003.819861