Noname manuscript No.

(will be inserted by the editor)

Duality-Based Algorithms for Total-Variation-

Regularized Image Restoration

Mingqiang Zhu · Stephen J. Wright ·

Tony F. Chan

Received: date / Accepted: date

Abstract Image restoration models based on total variation (TV) have be-
come popular since their introduction by Rudin, Osher, and Fatemi (ROF) in
1992. The dual formulation of this model has a quadratic objective with sep-
arable constraints, making projections onto the feasible set easy to compute.
This paper proposes application of gradient projection (GP) algorithms to the
dual formulation. We test variants of GP with different step length selection
and line search strategies, including techniques based on the Barzilai-Borwein
method. Global convergence can in some cases be proved by appealing to ex-
isting theory. We also propose a sequential quadratic programming (SQP) ap-
proach that takes account of the curvature of the boundary of the dual feasible
set. Computational experiments show that the proposed approaches perform
well in a wide range of applications and that some are significantly faster than
previously proposed methods, particularly when only modest accuracy in the
solution is required.
Keywords Image Denoising · Constrained Optimization · Gradient Projec-
tion
Computational and Applied Mathematics Report 08-33, UCLA, October, 2008 (Revised).
This work was supported by National Science Foundation grants DMS-0610079, DMS-
0427689, CCF-0430504, CTS-0456694, and CNS-0540147, and Office of Naval Research grant
N00014-06-1-0345. The work was performed while T. Chan was on leave at the National Sci-
ence Foundation as Assistant Director of Mathematics and Physical Sciences.

M. Zhu
Mathematics Department, UCLA, Box 951555, Los Angeles, CA 90095-1555, USA
E-mail: [email protected]
S. J. Wright
Department of Computer Sciences, University of Wisconsin, 1210 W. Dayton Street, Madi-
son, WI 53705, USA
E-mail: [email protected]
T. F. Chan
Mathematics Department, UCLA, Box 951555, Los Angeles, CA 90095-1555, USA
E-mail: [email protected]
2

1 Introduction

1.1 Background

Variational models have been extremely successful in a wide variety of image

restoration problems and remain one of the most active areas of research in
mathematical image processing and computer vision. The most fundamental
image restoration problem is perhaps denoising. It forms a significant prelim-
inary step in many machine vision tasks such as object detection and recog-
nition. Total variation (TV)-based image restoration models were first intro-
duced by Rudin, Osher, and Fatemi (ROF) in their pioneering work [17]. It
was designed with the explicit goal of preserving sharp discontinuities (edges)
in an image while removing noise and other unwanted fine-scale detail. ROF
formulated the following minimization problem:
Z
min |∇u| s.t. ku − f k22 ≤ σ 2 . (1)
u Ω

Here, Ω denotes the image domain, which will be taken to be a bounded

domain in Rn with Lipschitz boundary. Usually Ω is simply a rectangle in R2 ,
modeling the computer screen. The function f : Ω → R represents the given
observed image and σ 2 is an estimate of the variance of the noise in the image
f . The notation | · | represents the Euclidean (`2 ) norm on R2 . The objective
function in the formulation (1) is the TV semi-norm of u.
Rather than solving the constrained minimization problem (1), ROF and
subsequent researchers also formulated an unconstrained minimization prob-
lem which uses the TV term as a Tikhonov regularization:
Z
λ
min P (u) := |∇u| dx + ku − f k22 . (2)
u Ω 2

This above problem yields the same solution as (1) for a suitable choice of the
Lagrange multiplier λ (see [6]).
Recently, many researchers have proposed algorithms that make use of the
dual formulation of the ROF model; see, for example [8], [4], and [5]. To derive
this form, we first notice the TV semi-norm has the following equivalent forms
Z Z Z
|∇u| = 1
max ∇u · w = max −u∇ · w, (3)
Ω w∈C0 (Ω), |w|≤1 Ω |w|≤1 Ω

where w : Ω → R2 . The rightmost definition of the TV semi-norm is more

general since it requires the function u only to have bounded variation (BV),
not necessarily to be smooth. In fact, this is the formal definition of TV semi-
norm for the space of BV functions.
With this definition of TV, the ROF model becomes
Z
λ
min max −u∇ · w + ku − f k22 ,
u w∈C01 (Ω), |w|≤1 Ω 2
 3

where u and w are the primal and dual variables, respectively. The min-max
theorem (see e.g., [12, Chapter VI, Proposition 2.4]) allows us to interchange
the min and max, to obtain
Z
λ
max min −u∇ · w + ku − f k22 .
w∈C0 (Ω), |w|≤1 u
1
Ω 2

The inner minimization problem can be solved exactly as follows:

1
u=f+ ∇·w (4)
λ
leading to the following dual formulation:
" 2 #
λ 2
1
max D(w) := kf k2 − ∇ · w + f
, (5)
1
w∈C0 (Ω), |w|≤1 2 λ 2

or, equivalently,
1
min k∇ · w + λf k22 . (6)
w∈C01 (Ω), |w|≤1 2
For a primal-dual feasible pair (u, w), the duality gap G(u, w) is defined to
be the difference between the primal and the dual objectives:

G(u, w) = P (u) − D(w)

Z " 2 #
λ 2 λ 2
1
= |∇u| + ||u − f ||2 − kf k2 − ∇ · w + f

Ω 2 2 λ
2
Z   λ 2
1
= |∇u| − ∇u · w + ∇ · w + f − u
. (7)
Ω 2 λ

2

The duality gap bounds the distance to optimality of the primal and dual
objectives. Specifically, if u and w are feasible for the primal (2)) and dual (5)
problems, respectively, we have

0 ≤ P (u) − O∗ ≤ G(u, w), (8a)

∗
0 ≤ O − D(w) ≤ G(u, w), (8b)

where O∗ is the (common) primal-dual optimal objective value. In the dual-

based algorithms proposed in this paper, the primal variable u is calculated
using equation (4). This choice of u eliminates the second term in (7). We
make use of the duality gap in the termination test of Section 4.
Over the years, the ROF model has been extended to many other image
restoration tasks and has been modified in a variety of ways to improve its
performance (see [7] and the references therein). However, in our paper, we
will focus on the original TVL2 model (2) and more particularly on its dual
formulation (6).
4

1.2 Notation and Discrete Formulation

Before describing the numerical algorithms, let us fix our main notational
conventions.
Often in this paper we need to concatenate vectors and matrices, in both
column-wise or row-wise fashion. We follow the MATLAB convention of using
“,” for adjoining vectors and matrices in a row, and “;” for adjoining them in
a column. Thus, for any vectors x, y and z, the following are synonymous:
 
x
 y  = (xT , y T , z T )T = (x; y; z)
z

For simplicity, we assume that the domain Ω is the unit square [0, 1]×[0, 1],
and define a discretization via a regular n × n grid of pixels, indexed as (i, j),
for i = 1, 2, . . . , n, j = 1, 2, . . . , n. The index (i, j) represents the point (i/(n +
1), j/(n + 1)) ∈ Ω. We represent images images as two-dimensional matrices
of dimension n × n, where ui,j represents the value of the function u at the
point indexed by (i, j). (Adaptation to less regular domains is not difficult
in principle.) To define the discrete total variation, we introduce a discrete
gradient operator, whose two components at each pixel (i, j) are defined as
follows:

1 ui+1,j − ui,j if i < n
(∇u)i,j = (9a)
0 if i = n

ui,j+1 − ui,j if j < n
(∇u)2i,j = (9b)
0 if j = n.

(Thus ∇u ∈ Rn×n×2 .) The discrete TV of u is then defined by

X
TV(u) = k(∇u)i,j k.
1≤i,j,≤n

Here and throughout the paper, we use k · k and k · k2 interchangeably, to

denote the Euclidean (`2 ) norm of a vector of real numbers. Note that this
norm is not a smooth function of its argument. It has the classic “ice-cream
cone” shape, nondifferentiable when its argument vector is zero.
The discrete divergence operator is defined, by analogy with the continuous
setting, as the negative adjoint of the gradient operator, that is, ∇· = −∇∗ .
Defining the inner product of two objects in Rn×n as follows:
X
hu, vi = ui,j vi,j ,
1≤i,j≤n

(and similarly for objects in Rn×n×2 ), we have from definition of the discrete
divergence operator that for any u ∈ Rn×n and w ∈ Rn×n×2 , that h∇u, wi =
 5

hu, −∇ · wi. It is easy to check that the divergence operator can be defined
explicitly as follows:
 1 1
 2 2
 wi,j − wi−1,j if 1 < i < n  wi,j − wi,j−1 if 1 < j < n
1 2
(∇ · w)i,j = wi,j if i = 1 + w if j = 1
1  i,j2
−wi−1,j if i = n −wi,j−1 if j = n.


To describe the problem in matrix algebra language, we reorder the image

matrix u (resp. f ) in row-wise fashion into a vector v (resp. g), associating the
(i, j) element of the two-dimensional structure with the element (j − 1)n + i
of the vector structure, as follows:

v(j−1)n+i = ui,j , 1 ≤ i, j ≤ n.

We have v ∈ RN , where N = n2 . The (i, j) component of the gradient (9)

can thus be represented as a multiplication of the vector v ∈ RN by a matrix
ATl ∈ R2×N , for l = 1, 2, . . . , N :


(vl+1 − vl ; vl+n − vl ) if l mod n 6= 0 and l+n≤N

l+n − vl ) l+n≤N
(0; v if l mod n=0 and
ATl v = (10)
(vl+1 − vl ; 0)
 if l mod n 6= 0 and l+n>N

(0; 0) if l mod n=0 and l + n > N.


Using this notation, the discrete version of the primal ROF model (2) can be
written as follows:
N
X λ
min kATl vk2 + kv − gk22 (11)
v 2
l=1

Similarly, we restructure the dual variable w, using a row-wise ordering

of the indices (i, j), into a collection of vectors xl ∈ R2 , l = 1, 2, . . . , N , as
follows:  1 
wi,j
x(j−1)n+i = 2 , 1 ≤ i, j ≤ n.
wi,j

The complete vector x ∈ R2N of unknowns for the discretized dual problem
is then obtained by concatenating these subvectors: x = (x1 ; x2 ; . . . ; xN ). We
also form the matrix A by concatenating the matrices Al , l = 1, 2, . . . , N
defined in (10), that is, A = (A1 , . . . , AN ) ∈ RN ×2N . In this notation, the
divergence ∇ · w is simply −Ax, so the discretization of the dual ROF model
(6) is

1
min kAx − λgk22 (12)
2
x∈X

where X := {(x1 ; x2 ; . . . ; xN ) ∈ R2N : xl ∈ R2 ,

kxl k2 ≤ 1 for all l = 1, 2, . . . , N }.
6

1.3 A Fundamental Convergence Result

Here we make several remarks on the discretized problems (11), (12) and prove
a general convergence result. It is easy to verify that both problems can be
obtained from the function ` : RN × X → R defined as follows:
λ
`(v, x) := xT AT v + kv − gk22 . (13)
2
The primal problem (11) is simply

min max `(v, x),

v∈RN x∈X

while the dual problem (12) is equivalent to

max min `(v, x).

x∈X v∈RN

It is easy to verify that the conditions (H1), (H2), (H3), and (H4) of [15,
pp. 333-334] are satisfied by this setting. Thus, it follows from [15, Chap-
ter VII, Theorem 4.3.1] that ` has a nonempty compact convex set of saddle
points (v̄, x̄) ∈ RN × X. Moreover, from [15, Chapter VII, Theorem 4.2.5], the
point (v̄, x̄) ∈ RN × X is a saddle point if and only if v̄ solves (11) and x̄ solves
(12).
Note that by strict convexity of the objective in (11), the solution v̄ of
(11) is in fact uniquely defined. For any saddle point (v̄, x̄), we have that
`(v̄, x̄) ≤ `(v, x̄) for all v ∈ RN , that is, v̄ is a minimizer of `(·, x̄). Thus, from
optimality conditions for `(·, x̄), the following relationship is satisfied for the
unique solution v̄ of (11) and for any solution x̄ of (12):

Ax̄ + λ(v̄ − g) = 0. (14)

By uniqueness of v̄, it follows that Ax̄ is constant for all solutions x̄ of (12).
The following general convergence result will be useful in our analysis of
algorithms in Section 2.
Proposition 1 Let {xk } be any sequence with xk ∈ X for all k = 1, 2, . . .
such that all accumulation points of {xk } are stationary points of (12). Then
the sequence {v k } defined by
1 k
vk = g − Ax (15)
λ
converges to the unique solution v̄ of (11).

Proof Note first that all stationary points of (12) are in fact (global) solutions
of (12), by convexity.
Suppose for contradiction that v k 6→ v̄. Then we can choose  > 0 and a
subsequence S such that kv k − v̄k2 ≥  for all k ∈ S. Since all xk belong to
the bounded set X, the sequence {xk } is bounded, so {v k } is bounded also.
In particular, the subsequence {v k }k∈S must have an accumulation point v̂,
 7

which must satisfy kv̂ − v̄k2 ≥  > 0. By restricting S if necessary, we can

assume that limk∈S v k = v̂. By boundedness of {xk }, we can further restrict
S to identify a point x̂ ∈ X such that limk∈S xk = x̂. By (15), we thus have

Ax̂ + λ(v̂ − g) = 0 = lim Axk + λ(v k − g) = 0, (16)

k∈S

Since x̂ is an accumulation point of the whole sequence {xk }, we have by

assumption that x̂ is a stationary point and hence a solution of (12). By our
observation following (14), we thus have that Ax̂ + λ(v̄ − g) = 0, where v̄ is
the unique solution of (11). By comparing this expression with (16), we obtain
the contradiction v̂ = v̄, proving the result.

1.4 Previous Algorithms

We briefly review here some of the many algorithms that have been proposed
for solving the primal formulation (2) of the ROF model, the dual formulation
(6), or both formulations simultaneously. We refer the interested readers to [9]
for a more comprehensive survey.
In their original paper [17], ROF proposed a time-marching scheme that
solves the associated Euler-Lagrange equation of (2) by seeking the steady-
state solution of a parabolic PDE. The method is (asymptotically) slow due
to the CFL stability constraints (see [16]), which puts a tight bound on the
time step when the solution develop flat regions (where |∇u| ≈ 0). Hence, this
scheme is useful in practice only when low-accuracy solutions suffice. Even for
an accuracy sufficient to yield a visually satisfactory result, the cost is often
too great.
In [19], Vogel and Oman proposed to solve the same Euler-Lagrange equa-
tion of (2) via fixed-point iteration. Their main idea is to fix the diffusion
1
coefficient |∇u| in the Euler-Lagrange equation to its value at a previous step,
thus obtaining the solution to the nonlinear equation by solving a sequence of
linear systems. They prove global convergence and show that their method is
asymptotically much faster than the explicit time-marching scheme.
Chan, Golub, and Mulet [8] (CGM) use Newton’s method to solve a smoothed
version of the primal-dual system for the ROF model, in which p the gradient
norm |∇u| is replaced by a smoothed approximation |∇u|β = |∇u|2 + β,
for some smoothing parameter β > 0. Since this approach is based on New-
ton’s method, it converges quadratically, but to a solution of the smoothed
approximate model rather than to the solution of (2). Smaller values of β
yield approximate solutions that are closer to the true solution, but more iter-
ations are required before the onset of asymptotic quadratic convergence. The
cost per iteration is similar to that of the fixed-point iteration scheme.
Hintermüller and Stadler [14] (HS) discuss an
R infeasible-interior-point method
for a modification of (2) in which an term µ Ω |∇u|2 dx is added, for some
small but positive µ. By perturbing the dual of their problem with a regular-
ization term, then applying a semismooth Newton method to the primal-dual
8

formulation of the resulting problem, they obtain superlinear convergence to

a solution of the modified problem. In the implementation, the linear system
at each iteration is symmetrized and solved with iterative approaches, such
as preconditioned conjugate gradient. The overall approach is related to the
CGM method of [8] in that both methods use a Newton-like scheme to solve a
perturbation of the primal-dual optimality conditions. One difference is that
the HS method does not require the dual variable w to satisfy the constraints
|w| ≤ 1 strictly at every iteration, thus allowing longer steps along the Newton
direction to be taken at some iterations.
Goldfarb and Yin [13] reformulate the original ROF model (1) as a second-
order cone program (SOCP) and solve it with a standard primal-dual interior-
point software package (MOSEK). They replace MOSEK’s default ordering of
the variables with a nested dissection reordering strategy that is better tuned
to their formulation. In contrast to the method of [8], the SOCP formulation
converges to the true solution of the underlying ROF model (1).
There are also algorithms that tackle the dual formulation (6) explicitly.
Chambolle’s method [5] is the best known of this type. He invented the dual
semi-implicit gradient descent algorithm based on an original observation he
made concerning associated Lagrange multipliers. The method is globally con-
vergent, with suitable restriction on the time step, and is much faster than the
primal time-marching scheme.

1.5 Motivations and Proposed Approaches

Most existing numerical algorithms to solve ROF models (2) or (6) can be
loosely divided into two categories: those that need to solve a linear system
of equations at each iteration (implicit) and those that require only a matrix-
vector multiplication in the discrete setting (explicit). Generally speaking, the
implicit methods (e.g. CGM, HS, and SOCP) have fast asymptotic conver-
gence rates and can provide highly accurate benchmark solutions. However,
explicit methods are preferred in many situations for their simplicity and their
convergence with relatively little computational effort to medium-accurate and
visually satisfactory results. Their low memory requirements make them even
more attractive for large-scale problems. To illustrate the high memory require-
ments of implicit schemes, we note that an image of size 512 × 512 is close to
the limit of what the SOCP solver MOSEK can handle on a workstation with
2GB of memory.
In the remainder of this paper, we report on the development, implementa-
tion, and testing of some simple but fast explicit algorithms. These algorithms
are based on the the dual formulation (6) so they do not require any numerical
smoothing parameters that would prevent them from converging to the true
optimizer. Our proposed approaches are for the most part gradient projection
algorithms applied to (6), in which the search path from each iterate is ob-
tained by projecting negative-gradient (steepest descent) directions onto the
feasible set. Various enhancements involving different step-length rules and
 9

different line-search strategies are important in making the method efficient.

We also propose a sequential quadratic programming approach in which the
curvature of the boundary of the feasible set is taken into account.
For the general problem of optimizing a smooth function over a closed
convex set, that is,
min F (x) (17)
x∈X

(where F : R → R is smooth and X is a closed convex subset of Rm ),

m

gradient projection methods set

xk+1 = xk + γk (xk (αk ) − xk ), (18)
for some parameter γk ∈ [0, 1], where
xk (αk ) := PX (xk − αk ∇F (xk )), (19)
for some αk > 0. Here, PX denotes the projection onto the set X. Since X is
closed and convex, the operator PX is uniquely defined, but in order for the
gradient projection approach to make practical sense, this operator must also
be easy to compute. For this reason, gradient projection approaches have been
applied most often to problems with separable constraints, where X can be
expressed as a Cartesian product of low-dimensional sets. In our case (6), X is
a cross product of unit balls in R2 , so computation of PX requires only O(N )
operations. Bertsekas [2] gives extensive background on gradient projection
algorithms.

2 Gradient Projection Algorithms

From now on, we will focus on the solution of problem (12), which we restate
here:
1
min F (x) := kAx − λgk22 , (20)
x∈X 2
where the compact set X ⊂ R2N is defined in (12). In this section, we discuss
GP techniques for solving this problem. Our approaches move from iterate
xk to the next iterate xk+1 using the scheme (18)-(19). Projection PX on
the set X, a Cartesian product of unit Euclidean balls, can be computed
straightforwardly as follows.
  xl
PX (x) = , l = 1, 2, . . . , N. (21)
l max{kxl k, 1}
This operation projects each 2 × 1 subvector of x separately onto the unit
ball in R2 . It is worth pointing out here that this structure of the dual con-
straints, which makes the gradient projection approach practical, also enables
Chambolle to develop an analytical formula for the Lagrange multipliers in
[5].
Our approaches below differ in their rules for choosing the step-length
parameters αk and γk in (18) and (19).
10

2.1 Three Frameworks

We next consider three gradient projection frameworks that encompass our

gradient projection algorithms, and present convergence results for methods
in these frameworks.
Framework GP-NoLS (short for “gradient projection with no line search”)
chooses αk in some predetermined range and sets γk ≡ 1.

Framework GP-NoLS
Step 0. Initialization. Choose parameters αmin , αmax with 0 < αmin < αmax .
Choose x0 and set k ← 0.
Step 1. Choose step length αk ∈ [αmin , αmax ].
Step 2. Set xk+1 = xk (αk ).
Step 3. Terminate if a stopping criterion is satisfied; otherwise set k ← k + 1
and go to Step 1.

Framework GP-ProjArc also sets γk ≡ 1, but chooses αk by a backtracking

line search to satisfy a sufficient decrease condition relative to some reference
value of the objective. If the reference value is taken to be F (xk ), then this
approach yields monotonic decrease of F , and is referred to by Bertsekas [2,
p. 236] as Armijo Rule Along the Projection Arc.

Framework GP-ProjArc
Step 0. Initialization. Choose parameters αmin , αmax with 0 < αmin < αmax ,
and choose ρ ∈ (0, 1) and µ ∈ (0, 12 ). Choose x0 and set k ← 0.
Step 1. Choose initial step length ᾱk ∈ [αmin , αmax ].
Step 2. Backtracking Line Search. Choose reference value Frk , and set m to
the smallest nonnegative integer such that
F (xk (ρm ᾱk )) ≤ Frk − µ∇F (xk )T (xk − xk (ρm ᾱk )),
where xk (α) is defined as in (19);
Set αk = ρm ᾱk and xk+1 = xk (αk ).
Step 3. Terminate if a stopping criterion is satisfied; otherwise set k ← k + 1
and go to Step 1.

If we set Frk = +∞ for all k, then Framework GP-ProjArc reduces to

Framework GP-NoLS . An interesting choice for reference value is to set
(
max(F (xk ), F (xk−1 ), . . . , F (xk−M )) if k ≥ M ,
Frk := (22)
+∞ otherwise,
that is, the largest function value over the last M + 1 iterations, where M is
a small nonnegative integer (see [3]). Such a choice allows (limited) nonmono-
tonicity yet enables convergence results to be proved.
 11

Framework GP-LimMin fixes αk at the start of each iteration (possibly

using information gathered on previous steps), but then performs a “limited
minimization” procedure to find γk , again ensuring decrease of F at every
step.

Framework GP-LimMin
Step 0. Initialization. Choose parameters αmin , αmax with 0 < αmin < αmax .
Choose x0 and set k ← 0.
Step 1. Choose step length αk ∈ [αmin , αmax ]. Compute xk (αk ) and set δ k :=
(xk (αk ) − xk ).
Step 2. Limited Minimizing Line Search. Set xk+1 = xk + γk δ k , with γk =
mid(0, γk,opt , 1) and

−(δ k )T ∇F (xk )
γk,opt = arg min F (xk + γδ k ) = (23)
kAδ k k22

Step 3. Terminate if a stopping criterion is satisfied; otherwise set k ← k + 1

and go to Step 1.

The first algorithm we consider — Algorithm GPCL — is obtained from

Framework GP-NoLS by setting αk equal to the fixed value α > 0 at every
step. Convergence is obtained for all α sufficiently small, as we now show.
Theorem 1 Let {xk } be a sequence generated by Algorithm GPCL. Then if
0 < α < .25, the sequence v k obtained from (15) converges to the unique
solution v̄ of (11).

Proof Given any two vectors x0 and x00 we have that

∇F (x0 ) − ∇F (x00 ) = AT A(x0 − x00 ),

so the Lipschitz constant for ∇F is kAT Ak2 , which is bounded by 8 (see [5,
p. 92]). It follows immediately from [2, Proposition 2.3.2] that every accumu-
lation point of {xk } is stationary for (20) provided that 0 < α < .25. The
result now follows immediately from Proposition 1.

The upper bound of .25 in Theorem 1 is tight; we observe in practice that

the method is unstable even for α = .251.
For algorithms in Framework GP-ProjArc , we have the following conver-
gence result.
Theorem 2 Let {xk } be a sequence generated by an algorithm in Framework
GP-ProjArc , where we set Frk according to (22), for some nonnegative integer
M . Then the sequence v k obtained from (15) converges to the unique solution
v̄ of (11).
12

Proof In the case M = 0 (monotonic decrease), Proposition 2.3.3 of Bert-

sekas [2], with minor modifications for the variable choice of ᾱk within the
range [αmin , αmax ], shows that all accumulation points of {xk } are station-
ary. For general M , Theorem 2.3 of Birgin, Martı́nez, and Raydan [3] shows
likewise that all accumulation points of {xk } are stationary. (The latter proof
concerns an algorithm where ᾱk is chosen by a Barzilai-Borwein formula, but
in fact the proof holds for any choice of ᾱk in the range [αmin , αmax ].)
In both cases, we obtain the result from Proposition 1.
An identical result holds for algorithms in Framework GP-LimMin .
Theorem 3 Let {xk } be a sequence generated by an algorithm in Framework
GP-LimMin . Then the sequence v k obtained from (15) converges to the unique
solution v̄ of (11).
Proof Proposition 2.3.1 of Bertsekas [2], with minor modifications for the vari-
able choice of ᾱk within the range [αmin , αmax ], shows that all limit points of
{xk } are stationary. The result then follows from Proposition 1.

2.2 Barzilai-Borwein Strategies

We discuss strategies that choose αk using approaches first proposed by Barzi-

lai and Borwein [1] (BB) and subsequently elaborated by other authors. For
the unconstrained problem min F (x), the basic BB strategy sets xk+1 ←
xk − αk ∇F (xk ), where αk is chosen so that αk−1 I mimics the behavior of
the Hessian ∇2 F over the previous step. By Taylor’s theorem, we have
∇2 F (xk )∆xk−1 ≈ ∆g k−1 , ∆xk−1 ≈ (∇2 F (xk ))−1 ∆g k−1 ,
where
∆xk−1 := xk − xk−1 , ∆g k−1 := ∇F (xk ) − ∇F (xk−1 ),
so our desired property on α is that α−1 ∆xk−1 ≈ ∆g k−1 . Note that for the F
we consider here (20), we have ∆g k−1 = AT A∆xk−1 .
One formula for α is obtained by performing a least-squares fit in one
variable, as follows:
 −1
k−1 k−1 2
αk,1 = arg min kτ ∆x − ∆g k2 ,
τ ∈R

which yields
k∆xk−1 k22 k∆xk−1 k22
αk,1 = = . (24)
h∆xk−1 , ∆g k−1 i kA∆xk−1 k22
An alternative formula is obtained similarly, by doing a least-squares fit to α
rather than α−1 , to obtain
h∆xk−1 , ∆g k−1 i kA∆xk−1 k22
αk,2 = arg min k∆xk−1 − α∆g k−1 k22 = = .
α∈R k∆g k−1 k22 kAT A∆xk−1 k22
(25)
 13

These step lengths were shown in [1] to be effective on simple problems; a

partial analysis explaining the behavior was given. Numerous variants have
been proposed recently, and subjected to with theoretical and computational
analysis. The BB step-length rules have also been extended to constrained
optimization, particularly to bound-constrained quadratic programming; see,
for example [10] and [18]. The same formulae (24) and (25) can be used in
these cases to determine the step length, but we obtain xk+1 by projecting
xk − αk ∇F (xk ) onto the feasible set X, and possibly performing additional
backtracking or line-search modifications to ensure descent in F .
Other variants of Barzilai-Borwein schemes have been proposed in various
contexts. The cyclic Barzilai-Borwein (CBB) method proves to have better
performance than the standard BB in many cases (see for example [11] and
the references therein). In this approach, we recalculate the BB step length
from one of the formulae (24) or (25) at only every mth iteration, for some
integer m. At intervening steps, we simply use the last calculated value of αk .
There are alternating Barzilai-Borwein (ABB) schemes that switch between
the definitions (24) and (25), either adaptively or by following a fixed schedule.
Barzilai-Borwein techniques have been used previously to solve nonnegative
image restoration problems, in which the optimization formulation is a bound-
constrained quadratic program [20] rather than a TV-regularized least-squares
problem.

2.3 Implemented Variants of Gradient Projection

We discuss here the variants of gradient projection that were implemented in

our computational testing.

Algorithm GPLS. This algorithm falls into Framework GP-ProjArc , where

we choose the initial step length ᾱk at each iteration by predicting what the
step length would be if no new constraints were to become active on this step.
Specifically, we define the vector g k by
(
k (∇F (xk ))l , if kxkl k2 < 1 or (∇F (xk ))Tl xkl > 0,
gi =  k k T
 k
I − xl (xl ) (∇F (x ))l , otherwise.

We then choose the initial guess to be

ᾱk = arg min F (xk − αg k ),

α

which can be computed explicitly as

(g k )T ∇F (xk ) kg k k22
ᾱk = 2 = .
k
kAg k2 kAg k k22

In practice, we find that using 12 ᾱk as the initial value gives better performance,
and backtracking is not necessary in any of our numerical experiments.
14

Algorithm GPBB-NM. This is a nonmonotone Barzilai-Borwein method in

Framework GP-NoLS , in which we obtain the step length αk via the formula
(24), projected if necessary onto the interval [αmin , αmax ].

Algorithm GPBB-NM(m). A nonmonotone cyclic Barzilai-Borwein algorithm

in Framework GP-NoLS , in which αk is recalculated from (24) at every mth
iteration. Formally, we set

BB
αml+i = αml+1 for l = 0, 1, 2, . . . and i = 1, 2, . . . , m − 1,

BB
where αml+1 is obtained from (24) with k = ml + 1, restricted to the interval
[αmin , αmax ].

Algorithm GPBB-M. A monotone Barzilai-Borwein method in Framework

GP-LimMin , in which αk is obtained as in Algorithm GPBB-NM.

Algorithm GPBB-M(m). A monotone cyclic Barzilai-Borwein algorithm in

Framework GP-LimMin , in which αk is recalculated from (24) at every mth
iteration, similarly to Algorithm GPBB-NM(m).

Algorithm GPBB(safe). A nonmonotone Barzilai-Borwein algorithm in Frame-

work GP-ProjArc , in which ᾱk is calculated from (24) and Frk is defined as
in (22).

Algorithm GPABB. A monotonic alternating Barzilai-Borwein method in Frame-

work GP-LimMin , in which the technique of Serafini, Zanghirati, and Zanni [18,
Section 2.2], is used to switch between the rules (24) and (25). This technique
makes use of two positive integer parameters nmin and nmax with 0 < nmin ≤
nmax . Let nα be the number of consecutive iterations that use the same step-
length selection rule, (24) or (25). We switch from one rule to the other at the
next iteration k + 1 if either (i) nα ≥ nmax or (ii) nα ≥ nmin and αk is either a
separating step length or a bad descent generator. The current step length αk
is a separating step length if it lies between the values generated by the two
rules at the next iteration, that is, αk+1,2 < αk < αk+1,1 . Given two constants
γl and γu with 0 < γl ≤ 1 ≤ γu , we say that αk is a bad descent generator if
one of the following conditions holds:

(a) γk,opt < γl and αk = αk,1 ; or

(b) γk,opt > γu and αk = αk,2 .

where γk,opt is obtained from the limited minimization rule (23). We refer
interested readers to [18] for the rationale of the criterion. In any case, the
chosen αk is adjusted to ensure that it lies in the interval [αmin , αmax ].
 15

3 A Sequential Quadratic Programming Algorithm

We describe here a variation on the techniques of the previous section in

which the curvature of the boundary of the constraint set X is accounted for
in computing the search direction. The method can be viewed as a sequential
quadratic programming (SQP) method applied to the dual formulation (20).
The KKT optimality conditions for this formulation can be written as follows:

ATl (Ax − λg) + 2zl xl = 0, l = 1, 2, . . . , N,

2
0 ≤ zl ⊥ kxl k − 1 ≤ 0, l = 1, 2, . . . , N,

where the scalars zl are Lagrange multipliers for the constraints kxl k22 ≤ 1,
l = 1, 2, . . . , N , and the operator ⊥ indicates that at least one of its two
operands must be zero. At iteration k, we compute an estimate of the active
set Ak ⊂ {1, 2, . . . , N }, which are those indices for which we believe that
kxl k22 = 1 at the solution. In our implementation, we choose this set as follows:

Ak = {l | kxkl k2 = 1 and (xkl )T [∇F (xk )]l ≤ 0}

= {l | kxkl k2 = 1 and (xkl )T ATl (Axk − λg) ≤ 0}. (26)

The SQP step is a Newton-like step for the following system of nonlinear
equations, from the current estimates xk and zlk , l = 1, 2, . . . , N :

ATl (Ax − λg) + 2xl zl = 0, l = 1, 2, . . . , N, (27a)

kxl k22 − 1 = 0, l ∈ Ak , (27b)
zl = 0, l∈
/ Ak . (27c)

Using z̃lk+1 to denote the values of zl at the next iterate, and d˜k to denote
the step in xk , a “second-order” step can be obtained from (27) by solving the
following system for d˜k and z̃lk+1 , l = 1, 2, . . . , N :

ATl Ad˜k + 2z̃lk+1 d˜kl = −ATl [Axk − λg] − 2xkl z̃lk+1 , l = 1, 2, . . . , N, (28a)
2(xkl )T d˜kl = 0, l ∈ Ak , (28b)
z̃lk+1 = 0, l ∈
/ Ak . (28c)

We now define Newton-like steps dk in x, and new iterates z k+1 in z, by

replacing AT A by αk−1 I in (28a) and solving the following linear system:

αk−1 dkl + 2zlk+1 dkl = −ATl [Axk − λg] − 2xkl zlk+1 , l = 1, 2, . . . , N, (29a)
2(xkl )T dkl = 0, l ∈ Ak , (29b)
zlk+1 = 0, l ∈
/ Ak . (29c)

Considering indices l ∈ Ak , we take the inner product of (29a) with xkl and
use (29b) and (26) to obtain:

zlk+1 = −(1/2)(xkl )T ATl (Axk − λg), l ∈ Ak .

16

We obtain the steps dkl for these indices by substituting this expression in
(29a):
dkl = −(αk−1 + 2zlk+1 )−1 ATl (Axk − λg) + 2xkl zlk+1 , l ∈ Ak .
 

In fact, because of (29c), this same formula holds for l ∈

/ Ak , when it reduces
to the usual negative-gradient step
dkl = −αk ATl (Axk − λg), l ∈
/ Ak .
We define the (nonmonotone) Algorithm SQPBB-NM by making an initial
choice of αk at each iteration according to the formula (24), and calculating
xk+1 = xk + dk and z k+1 as described above. In the monotone variant of
this method, known as SQPBB-M, we successively decrease αk by a factor of
ρ ∈ (0, 1), as in Framework GP-ProjArc , and recalculate xk+1 and z k+1 as
above, until a decrease in the objective function is obtained.
We also tried versions of these methods in which αk was recalculated only
on every mth iteration; these are referred to as SQPBB-NM(m) and SQPBB-
M(m), respectively.

4 Termination

The decision about when an approximate solution is of sufficiently high quality

to terminate the algorithm can be difficult for general constrained optimization
problems. Often, we wish the approximate solution x to be close to a global
minimizer x∗ and/or the function value F (x) be close to F (x∗ ). In the case
of (20), the duality gap (7) provides a reliable and easily calculated stopping
criterion.
If (u, w) be a feasible primal-dual pair satisfying (4), we have from (7) that
Z  
G(u, w) = |∇u| − ∇u · w . (30)
Ω

We terminate the algoithm when the current iterate w = wk satisfies the

following stopping criterion:
G(u, w)
≤ tol, (31)
|P (u)| + |D(w)|
where u is obtained from (4) and tol is a small positive tolerance. It follows
from (8b) that the dual objective D(w) is close to the optimal objective O∗
when (31) is satisfied, in the sense that
0 ≤ O∗ − D(w) ≤ (tol)(|P (u)| + |D(w)|).
We can show that the u obtained from (4) is also close to the optimal value
u∗ when this test is satisfied, by the following argument. From (4), we have
1 1 1
u − u∗ = (f + ∇ · w) − (f + ∇ · w∗ ) = (∇ · w − ∇ · w∗ )
λ λ λ
 17

Since |∇u∗ | = ∇u∗ · w∗ and |∇u| ≥ ∇u · w for any feasible w (since |w| ≤ 1),
we have
Z
λku − u∗ k22 = (u − u∗ )(∇ · w − ∇ · w∗ )
Ω
Z Z Z Z
∗ ∗
= u∇ · w − u∇ · w − u ∇·w+ u∗ ∇ · w ∗
Ω Ω Ω Ω
Z Z Z Z
∗ ∗
= −∇u · w + ∇u · w + ∇u · w − |∇u∗ |
ZΩ ZΩ Z Ω Ω
∗
≤ −∇u · w + |∇u| + |∇u |(|w| − 1)
ZΩ Ω Ω

≤ (|∇u| − ∇u · w)
Ω
= G(u, w). (32)

Using this bound, we obtain the following bound when (u, w) satisfies (31):
p p
ku − u∗ k2 ≤ G(u, w)/λ ≤ (|P (u)| + |D(w)|)(tol)/λ.

5 Computational Experiments

We report on computational experiments for three test problems in image

denoising. The original clean images and the noisy images used as input to
the denoising codes are shown in Figure 1. The sizes of the discretizations for
the three test problems are 128 × 128, 256 × 256, and 512 × 512, respectively.
The noisy images are generated by adding Gaussian noise to the clean images
using the MATLAB function imnoise, with variance parameter set to 0.01.
The fidelity parameter λ is taken to be 0.045 throughout the experiments.
This parameter is inversely related to the noise level σ and usually needs to
be tuned for each individual image to get an optimal visual result.
We tested the following algorithms:
– Chambolle’s semi-implicit gradient descent method [5];
– many variants of gradient projection proposed in Section 2;
– the SQP method of Section 3;
– the CGM method of [8].
We report on a subset of these tests here, including the gradient projection
variants that gave consistently good results across the three test problems.
In Chambolle’s method, we take the step length to be 0.248 for near-
optimal run time, although global convergence is proved in [5] only for step
lengths in the range (0, .125). We use the same value αk = 0.248 in Algorithm
GPCL, as in this case it is justified by Theorem 1 and also gives near-optimal
run time.
For all gradient projection variants, we set αmin = 10−5 and αmax = 105 .
(Performances are insensitive to these choices, as long as αmin is sufficiently
18

Fig. 1 The original clean images for our test problems (left) alongside noisy images gener-
ated by adding Gaussian noise with variance .01, by calling Matlab function imnoise. Top:
128 × 128 “shape”; middle: 256 × 256 “cameraman”; Bottom: 512 × 512 “Barbara”.
 19

small and αmax sufficiently large.) In Algorithm GPLS, we used ρ = 0.5 and
µ = 10−4 . In Algorithm GPABB, we set γl = 0.1 and γu = 5. In GPBB(safe),
we set µ = 10−4 and M = 5 in the formula (22).
We also tried variants of the GPBB methods in which the initial choice
of αk was scaled by a factor of 0.5 at every iteration. We found that this
variant often enhanced performance. This fact is not too surprising, as we
can see from Section 3 that the curvature of the boundary of constraint set X
suggests that it is appropriate to add positive diagonal elements to the Hessian
approximation, which corresponds to decreasing the value of αk .
In the CGM implementation, we used a direct solver for the linear system
at each iteration, as the conjugate gradient iterative solver (which is an option
in the CGM code) was slower on these examples. The smooth parameter β
is dynamically updated based on duality gap from iteration to iteration. In
2
particular, we take β0 = 100 and let βk = βk−1 (Gk /Gk−1 ) , where Gk and
Gk−1 are the duality gaps for the past two iterations. This simple strategy for
updating β, which is borrowed from interior-point methods, outperforms the
classical CGM approach, producing faster decrease in the duality gap.
All methods are coded in MATLAB and executed on a Dell Precision T5400
workstation with 2.66 GHz Inel Quadcore processor and 4GB main memory.
It is likely the performance can be improved by recoding the algorithms in C
or C++, but we believe that improvements would be fairly uniform across all
the algorithms.
Tables 1, 2, and 3 report number of iterations and average CPU times over
ten runs, where each run adds a different random noise vector to the true
image. In all codes, we used the starting point x0 = 0 in each algorithm and
the relative duality gap stopping criterion (31). We vary the threshold tol
from 10−2 to 10−6 , where smaller values of tol yield more accurate solutions
to the optimization formulation.
Figure 2 shows the denoised images obtained at different values of tol.
Note that there is little visual difference between the results obtained with
two tolerance values 10−2 and 10−4 . Smaller values of tol do not produce
further visual differences. By varing λ slightly we can obtain better visual
results for these images, but still the visual quality of the computed solution
does not improve markedly as tol is reduced below 10−2 .
The tables show that on all problems, the proposed gradient projection
algorithms are competitive to Chambolle’s method, and that some variants
are significantly faster, especially when moderate accuracy is required for
the solutions. Three variants stood out as good performers: the GPBB-NM
and GPABB variants, along with the GPBB-M(3) variant in which the initial
choice of αk was scaled by 0.5 at each iteration. For all tests with tol = 10−2 ,
tol = 10−3 , and tol = 10−4 , the winner was one of the gradient-projection
Barzilai-Borwein strategies.
For these low-to-moderate accuracy requirements, CGM is generally slower
than the gradient-based methods, particularly on the larger problems. The
picture changes somewhat, however, when high accuracy (tol = 10−6 ) is
20

Fig. 2 The denoised images with different level of termination criterions. left column: tol =
10−2 , right column: tol = 10−4 .
 21

Table 1 Number of iterations and CPU times (in seconds) for problem 1. ∗ = initial αk
scaled by 0.5 at each iteration.
tol = 10−2 tol = 10−3 tol = 10−4 tol = 10−6
Algorithms Iter Time Iter Time Iter Time Iter Time
Chambolle 18 0.06 164 0.50 1091 3.34 22975 74.26
GPCL 36 0.11 134 0.38 762 2.22 15410 46.34
GPLS 22 0.14 166 1.02 891 5.63 17248 113.84
GPBB-M 12 0.05 148 0.61 904 3.75 16952 72.82
GPBB-M(3) 13 0.05 69 0.28 332 1.32 4065 16.48
GPBB-M(3)∗ 11 0.05 47 0.18 188 0.75 2344 9.50
GPBB-NM 10 0.04 49 0.18 229 0.85 3865 14.66
GPABB 13 0.07 54 0.26 236 1.13 2250 10.90
GPBB(safe) 10 0.05 50 0.21 209 0.98 3447 17.32
SQPBB-M 11 0.07 48 0.32 170 1.20 3438 25.05
CGM 5 1.13 9 2.04 12 2.73 18 4.14

Table 2 Number of iterations and CPU times (in seconds) for problem 2. ∗ = initial αk
scaled by 0.5 at each iteration.
tol = 10−2 tol = 10−3 tol = 10−4 tol = 10−6
Algorithms Iter Time Iter Time Iter Time Iter Time
Chambolle 26 0.36 165 2.21 813 11.03 14154 193.08
GPCL 32 0.41 116 1.49 535 7.03 9990 132.77
GPLS 24 0.55 134 3.84 583 17.48 11070 341.37
GPBB-M 20 0.38 124 2.35 576 10.93 10644 203.83
GPBB-M(3) 20 0.36 98 1.78 333 6.09 3287 60.44
GPBB-M(3)∗ 17 0.31 47 0.86 167 3.05 1698 31.17
GPBB-NM 16 0.27 53 0.91 183 3.15 2527 43.80
GPABB 16 0.36 47 1.02 158 3.46 1634 35.85
GPBB(safe) 16 0.29 52 0.96 170 3.45 2294 51.87
SQPBB-M 14 0.37 48 1.36 169 5.16 2537 79.83
CGM 5 5.67 9 10.37 13 15.10 19 22.28

required. The rapid asymptotic convergence of CGM is seen to advantage in

this situation, and its runtime improves on all variants of gradient projection.
Figure 3 plots the relative duality gap (the left-hand side of (31) against
the CPU time cost for Chambolle’s method, CGM method, and the GPBB-
NM variant of the gradient projection algorithm for each of the three test
problems. Convergence of the “best so far” iterates for GPBB-NM (the lower
bound of this curve) is clearly faster than Chambolle’s method, and also faster
than CGM until a high-accuracy solution is required.

6 Conclusions and Final Remarks

We have proposed gradient projection algorithms for solving the discretized

dual formulation of the total variation image restoration model of Rudin, Os-
her, and Fatemi [17]. The problem has a convex quadratic objective with sep-
arable convex constraints, a problem structure that makes gradient projection
schemes practical and simple to implement. We tried different variants of gra-
dient projection (including non-monotone Barzilai-Borwein spectral variants)
22

6
10
BB−NM
Chambolle
CGM
4
10

2
10

Relative Duality Gap

0
10

−2
10

−4
10

−6
10

−8
10
0 5 10 15 20 25 30
CPU Time
8
10
BB−NM
Chambolle
6 CGM
10

4
10

2
10
Relative Duality Gap

0
10

−2
10

−4
10

−6
10

−8
10
0 10 20 30 40 50 60 70 80 90
CPU Time
8
10
BB−NM
Chambolle
6 CGM
10

4
10

2
10
Relative Duality Gap

0
10

−2
10

−4
10

−6
10

−8
10
0 50 100 150 200 250 300 350
CPU Time

Fig. 3 Duality gap vs. CPU time for GPBB-NM, Chambolle, and CGT codes, for problems
Prolems 1, 2, and 3, respectively.
 23

Table 3 Number of iterations and CPU times (in seconds) for problem 3. ∗ = initial αk
scaled by 0.5 at each iteration.
tol = 10−2 tol = 10−3 tol = 10−4 tol = 10−6
Algorithms Iter Time Iter Time Iter Time Iter Time
Chambolle 27 2.35 131 11.43 544 47.16 8664 751.59
GPCL 24 2.02 80 6.83 337 28.76 5907 505.86
GPLS 40 5.86 89 14.72 324 57.87 5780 1058.59
GPBB-M 20 2.44 88 10.82 352 43.33 5644 695.89
GPBB-M(3) 20 2.38 59 7.09 197 23.73 2598 313.59
GPBB-M(3)∗ 17 2.03 41 4.94 131 15.77 1191 143.66
GPBB-NM 15 1.67 40 4.58 117 13.20 1513 171.21
GPABB 14 1.99 38 5.41 117 16.76 1135 162.54
GPBB(safe) 15 1.74 39 4.70 112 14.61 1425 204.50
SQPBB-M 14 2.35 35 6.24 93 17.19 1650 312.16
CGM 6 42.38 10 67.65 13 87.66 19 126.68

that make use of different step-length strategies. We compare these methods

to two popular existing approaches proposed by Chambolle [5] and Chan,
Golub, and Mulet [8], and show that when low to moderate solution accuracy
is required, several of the gradient projection variants are faster than earlier
approaches.
Besides giving evidence of improved strategies for obtaining lower-accuracy
solutions to the ROF model, our results also suggest several strategies for
obtaining higher-accuracy solutions. First, it would seem appealing to use a
GP approach to obtain a low-accuracy solution, then “cross over” to CGM,
using the solution obtained from the GP approach to generate a starting point
for CGM. However, we found that the primal-dual starting point (v, x) = (g −
Ax/λ, x) obtained from the GP solution x was not generally better than using
a “cold start” for CGM. As in interior-point primal-dual methods, it seems
difficult to find good warm starts. Second, we experimented with a multi-level
strategy, using computed solutions for a coarser discretization (smaller n and
N ) to construct a starting point for a finer discretization. We have not yet
been able to find an interpolation strategy for which the constructed starting
point is better than a cold start.
The MATLAB implementations used to obtain the results of this paper,
along with data sets, are available at www.cs.wisc.edu/~swright/TVdenoising/.

Acknowledgements We thank the referees for several suggestions that improved the pa-
per.

References

1. Barzilai, J., Borwein, J.M.: Two-point step size gradient methods. IMA Journal of
Numerical Analysis 8, 141–148 (1988)
2. Bertsekas, D.P.: Nonlinear Programming, second edn. Athena Scientific (1999)
3. Birgin, E.G., Martı́nez, J.M., Raydan, M.: Nonmonotone spectral projected gradient
methods on convex sets. SIAM Journal on Optimization 10(4), 1196–1211 (2000)
24

4. Carter, J.L.: Dual method for total variation-based image restoration. Report 02-13,
UCLA CAM (2002)
5. Chambolle, A.: An algorithm for total variation minimization and applications. Journal
of Mathematical Imaging and Visualization 20, 89–97 (2004)
6. Chambolle, A., Lions, P.L.: Image recovery via total variation minimization and related
problmes. Numerische Mathematik 76, 167–188 (1997)
7. Chan, T.F., Esedoglu, S., Park, F., Yip, A.: Total variation image restoration: Overview
and recent developments. In: N. Paragios, Y. Chen, O. Faugeras (eds.) Handbook of
Mathematical Models in Computer Vision. Springer (2005)
8. Chan, T.F., Golub, G.H., Mulet, P.: A nonlinear primal-dual method for total variation
based image restoration. SIAM Journal of Scientific Computing 20, 1964–1977 (1999)
9. Chan, T.F., Zhu, M.: Fast algorithms for total variation-based image processing. In:
Proceedings of the 4th ICCM. Hangzhuo, China (2007)
10. Dai, Y.H., Fletcher, R.: Projected barzilai-borwein methods for large-scale box-
constrained quadratic programming. Numerische Mathematik 100, 21–47 (2005)
11. Dai, Y.H., Hager, W.W., Schittkowski, K., Zhang, H.: The cyclic Barzilai-Borwein
method for unconstrained optimization. IMA Journal of Numerical Analysis 26, 604–
627 (2006)
12. Ekeland, I., Témam: Convex Analysis and Variational Problems. SIAM Classics in
Applied Mathematics. SIAM (1999)
13. Goldfarb, D., Yin, W.: Second-order cone programming methods for total variation-
based image restoration. SIAM Journal on Scientific Computing 27, 622–645 (2005)
14. Hintermüller, M., Stadler, G.: An infeasible primal-dual algorithm for TV-based inf-
convolution-type image restoration. SIAM Journal on Scientific Computing 28, 1–23
(2006)
15. Hiriart-Urruty, J., Lemaréchal, C.: Convex Analysis and Minimization Algorithms,
vol. I. Springer-Verlag, Berlin (1993)
16. Osher, S., Marquina, A.: Explicit algorithms for a new time dependent model based on
level set motion for nonlinear deblurring and noise removal. SIAM Journal of Scientific
Computing 22, 387–405 (2000)
17. Rudin, L., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algo-
rithms. Physica D 60, 259–268 (1992)
18. Serafini, T., Zanghirati, G., Zanni, L.: Gradient projection methods for large quadratic
programs and applications in training support vector machines. Optimization Methods
and Software 20(2–3), 353–378 (2004)
19. Vogel, C.R., Oman, M.E.: Iterative methods for total variation denoising. SIAM Journal
of Scientific Computing 17, 227–238 (1996)
20. Wang, Y., Ma, S.: Projected barzilai-borwein methods for large-scale nonnegative image
restoration. Inverse Problems in Science and Engineering 15(6), 559–583 (2007)