Signal, Image and Video Processing

https://doi.org/10.1007/s11760-019-01415-6

ORIGINAL PAPER

Recursive SURE for image recovery via total variation minimization

Feng Xue1 · Jiaqi Liu1 · Xia Ai1

Received: 28 October 2017 / Revised: 31 October 2018 / Accepted: 3 January 2019

© Springer-Verlag London Ltd., part of Springer Nature 2019

Abstract
Recently, total variation regularization has become a standard technique, and even a basic tool for image denoising and
deconvolution. Generally, the recovery quality strongly depends on the regularization parameter. In this work, we develop a
recursive evaluation of Stein’s unbiased risk estimate (SURE) for the parameter selection, based on specific reconstruction
algorithms. It enables us to monitor the evolution of mean squared error (MSE) during the iterations. In particular, to deal
with large-scale data, we propose a Monte Carlo simulation for the practical computation of SURE, which is free of any
explicit matrix operation. Experimental results show that the proposed recursive SURE could lead to highly accurate estimate
of regularization parameter and nearly optimal restoration performance in terms of MSE.

Keywords Total variation · Denoising · Deconvolution · Stein’s unbiased risk estimate (SURE) · Jacobian recursion

1 Introduction 
N
2 2
TV(x) = (D1 x)n + (D2 x)n +α (3)
n=1
Problem statement—Consider the standard image recovery
problem: find a good estimate of original image x0 ∈ R N where D1 and D2 denote the horizontal and vertical first-order
from the following degradation model [15,20,22]: differences, respectively. The parameter α = 0 corresponds
to the standard TV definition. We use the smooth approx-
y = Hx0 +  (1)
imation with small α > 0, since it simplifies numerical
computations due to the differentiability [9].
where y ∈ R N is the observed image, H ∈ R N ×N denotes
TV is particularly effective for recovering those signals
the observation matrix, which represents either identity for
with piecewise constant region while preserving edges [12].
denoising or convolution for deconvolution, and  ∈ R N is an
Recently, people extended the basic TV-norm to more gen-
additive white Gaussian noise with known variance σ 2 > 0.
eral form of ϕ(Dx2 ) that models a priori of the first-order
Since the seminal work of ROF [12], total variation (TV)
gradients of an image [13]. Here, ϕ is a potential, possibly
regularization has become a standard technique [17,19]:
non-convex, function. It is reduced to the standard TV-norm
1  when ϕ(t) = t. In this work, we focus on the TV mini-
xλ = arg min
 Hx − y2 + λ · TV(x) (2)
2 mization and attempt to find a proper value of λ for a good
2
x
 
L(x)
restoration quality. This work may help to gain some insights
into more complicated function of ϕ.
Here, λ is a regularization parameter, which is essential for Related works—There have been a number of criteria for this
the recovery quality of 
xλ . The isotropic TV term is defined selection of λ, for example:
as [9]:
– Generalized cross-validation [5]: It is often used for linear
This work was supported by the National Natural Science Foundation
of China (No. 61401013). estimates, not applicable for the nonlinear reconstruction
considered here.
B Feng Xue – L-curve method [7]: This procedure is not fully auto-
[email protected] mated and often requires hand tuning or selection.
1 National Key Laboratory of Science and Technology on Test
Physics and Numerical Mathematics, Beijing 100076, China

123
 Signal, Image and Video Processing

– Discrepancy principle [8]: This criterion is easy to com- 2 Recursive evaluation of SURE for TV
pute and however may cause a loss of restoration quality.1 denoising

In this paper, we quantify the restoration performance by Now, we consider image denoising problem, i.e. H = I in
the mean squared error (MSE) [1,22]: (2). To perform the SURE-based selection of λ, we need to
compute the solution 
xλ and its SURE.
1  2
MSE = E xλ − x0 2 (4)
N 2.1 Basic scheme of Chambolle’s algorithm [2]
and attempt to select a value of λ, such that the corresponding
solution xλ achieves minimum MSE. Many algorithms can be used to find the TV solution xλ , e.g.
Notice that the MSE is inaccessible due to the unknown [3,4,6,11,14]. Here, we apply a dual-based iterative projec-
x0 . In practice, Stein’s unbiased risk estimate (SURE) has tion algorithm—Chambolle’s algorithm—to solve (2), since
been proposed as a statistical substitute for MSE [1,16]: it is one of the most popular TV minimization solvers and
has been extensively used in the recent decade. The original
1  2   form of the Chambolle’s iteration was described in [2]. Now,
SURE = 
xλ 2 − 2yT H−T
xλ + 2σ 2 Tr H−T Jy (
xλ )
N we rewrite the algorithm in matrix language:
1  2
+ x0 2 (5) ⎛
w(i)
⎞
N   
(i) ⎜ (i) τ ⎟
u(i+1) =V ⎜ ⎝u − λ D (y + λDT u(i) )⎟ (6)
since it depends on the observed data y only.2 Tr in (5) denotes   ⎠
xλ ) ∈ R N ×N is a Jacobian matrix
the matrix trace. Here, Jy ( x(i)
defined as [21,23]:
where τ is a step size and the gradient operator D is D =
  xλ )m
∂(
Jy (
xλ ) m,n = [DT1 , DT2 ]T ∈ R2N ×N [D1 and D2 are the same as in (3)].
∂ yn The vector u(i) ∈ R2N lives in D-domain of an image, which
can be transformed back to the image domain by DT u. The
The statistical unbiasedness of SURE w.r.t. true MSE has (i)
been proved in [22]. Recently, SURE has become a popular diagonal matrix V is:
criterion for parameter selection, in the context of nonlinear  (i) 
(i) V 0
denoising/deconvolution [1,22], and 1 -based sparse recov- V = ∈ R2N ×2N
ery [18,23]. However, to our best knowledge, there are very 0 V(i)
few researches on the application of SURE to TV-based
reconstruction, which is the purpose of this paper. where the diagonal block V(i) ∈ R N ×N is given by:

Our contributions—Our main contributions are twofold.  −1

(i) τ  (i)
2 
(i)
2
First, we develop a recursive evaluation of SURE during the Vn,n = 1+ (D1 x )n + (D2 x )n + α
λ
reconstruction iterations, which finally provides a reliable
estimate of the MSE for the TV-based recovery. Second, the
Monte Carlo (MC) simulation is used to facilitate the SURE
computation for large-scale data, without explicit matrix 2.2 Recursive evaluation of SURE
operation.
Additional remarks—Throughout this paper, we use bold- The next question is how to compute the SURE of the TV-
face lowercase letters, e.g. x ∈ R N , to denote N -dimensional denoised image  xλ ? We propose to compute the SURE of
real vectors, where N is typically the number of pixels in an x(i) during each iteration until final convergence. The similar
image. The matrices are denoted by boldface uppercase let- treatment has been used in [18,21,23] for 1 -based sparse
ters, e.g. A ∈ R M×N . AT ∈ R N ×M denotes the transpose of deconvolution.
matrix A. The superscripts (i) and ( j) denote the iterative The SURE for the ith update is (noting H = I for denois-
indices of outer or inner loops. The notation diag(v) trans- ing problem):
forms the vector v to the diagonal matrix V with Vn,n = vn . 
1  2
x(i) − y2 + 2σ Tr Jy (x(i) ) − σ 2
SURE = 2
(7)
1 See Sect. 4 for the complete comparisons between discrepancy prin- N N
ciple and the proposed SURE.
2Note that the last constant term—x0 22 /N —is irrelevant to the opti- The Jacobian matrix Jy (x(i) ) can be evaluated by the basic
mization of 
xλ . calculus as follows.

123
Signal, Image and Video Processing

First, we split (6) into two parts—vertical and horizontal Noting that x(i) = y + λDT u(i) in (6), we have:
differences:  
(i) (i)

(i+1)
    (i)  Jy (x(i) ) = I + λDT1 Jy u1 + λDT2 Jy u2 (11)
u1 V(i) 0 w1
(i+1) = 0 V (i) (i)
u2 w2
         Thus, the Jacobian matrix Jy (x(i) ) can be evaluated in this
(i)
u(i+1) V w(i) recursive manner, until the convergence of Chambolle’s iter-
ation, summarized in Algorithm 1.
(i+1)
From (6), the Jacobian matrix of u1 is:

  ∂V(i) (i) Algorithm 1: SURE evaluation for Chambolle’s denois-

(i+1) (i) m,m (i) ∂(w1 )m
Jy (u1 ) = w1 + Vm,m ing algorithm
m,n m ∂ yn ∂ yn
Input: y, λ, α, τ , initial u(0)
Output: reconstructed  xλ and SURE( xλ )
Let a = D1 x(i) and b = D2 x(i) , we have: for i = 1, 2, . . . (Chambolle’s iteration) do
1 compute x(i) by (6);
(i)
∂Vm,m ∂Vm,m ∂am
(i)
∂Vm,m ∂bm
(i) 2 update Jy (x(i) ) by (10) and (11);
= · + · 3 compute SURE of ith iterate by (7);
∂ yn ∂am ∂ yn ∂bm ∂ yn end
(i)
τ am (Vm,m )2 ∂am
=− ·! ·
λ a + b + α ∂ yn
2 2
 m  m  This algorithm enables us to solve the TV denoising prob-
(i)
(C1 )m,m lem with a prescribed value of λ and simultaneously evaluate
(i)
τ bm (Vm,m )2 ∂bm the SURE during the Chambolle’s iterations.
− ·! ·
λ a + b + α ∂ yn
2 2
 m  m  2.3 Monte Carlo for practical computation
(i)
(C2 )m,m
τ  (i)
=− C D1 Jy (x(i) ) − C(i) (i)
2 D2 Jy (x )
From (10)–(11), we can see that the Jacobian recursions
λ 1 m,n require the explicit matrix computations. However, for a typ-
ical image of size 256 × 256, the related matrices, e.g. Ws(i) ,
Thus, we obtain: (i)
Cs , V(i) and Ds (s = 1, 2), are of size 2562 × 2562 . Due to
 τ  the limited computational resources (e.g. RAM), it is imprac-
Jy u1(i+1) = − W1(i) C(i) (i) (i)
1 D1 + C2 D2 Jy (x ) tical to explicitly store and compute such the huge matrices.
λ
(i) Thus, the Jacobian recursions cannot be computed in the
+V(i) Jy (w1 ) (8)
explicit matrix form of (10)–(11).
(i) (i)
However, Monte Carlo simulation provides an alternative
where diagonal matrix W1 = diag(w1 ). way to compute the trace by the following fact [1]:
(i+1)
Similarly, Jy (u2 ) is given by:

 τ (i) (i)  Tr Jy (x(i) ) = E n0T Jy (x(i) )n0 (12)
Jy u2
(i+1) (i)
= − W2 C1 D1 + C2 D2 Jy (x(i) )   
λ  nx
(i)

+V(i) Jy w2(i) (9)

with input white Gaussian noise n0 ∼ N (0, I N ), provided
(i)
By the basic property of Jacobian matrix [23], we further that nx can be computed without explicit form of Jy (x(i) ).
(i) (i) Multiplying the input n0 on both sides of (10)–(11), we
have Jy (ws ) = Jy (us ) − τλ Ds Jy (x(i) ) for s = 1, 2. Sub-
stituting Jy (w(i) ) into Jy (u(i) ) yields (after rearrangements): obtain:
⎧
⎡ ⎤ ⎪
⎪
(i+1)
nu1
(i)
nu1 (i)
 (i)
(i+1)  ⎪
⎪       nx
Jy u1 V 0 τ ⎪
⎪     
Jy (u (i+1)
)= ⎣ = ⎦ (i) Jy (u(i) ) − · ⎪
⎨ Jy u(i+1) n0 = V(i) Jy u(i) n0 − τ P(i) Jy (x(i) )n0
(i+1)
0 V λ λ 1
Jy u2 1  1 
  ⎪ (i+1) (i) J (i) τ (i) (i)
(i) (i) (i) (i) ⎪
⎪ J y u n0 = V y u n0 λ P2 Jy (x )n0
−
W1 C1 + V(i) W1 C2 ⎪
⎪ 
2
  
2
    
DJy (x(i) ) ⎪
⎪
(i) (i)
W2 C1
(i) (i)
W2 C2 + V(i) ⎩ n
(i+1)
n
(i) n
(i)
x
u2 u2

(10) (13)

123
 Signal, Image and Video Processing

and convolution in (1)] [17,19]. By the variable splitting [17,19],

  (2) is equivalent to the following problem:
(i) (i)
Jy (x(i) )n0 = n0 + λDT
1 Jy u1 n0 +λDT
2 Jy u2 n0 (14)
        
(i)
nx (i) (i) 1 
Hx − y2 + λ · TV(z), s.t. z = x
nu1 nu2 min 2
x 2
(i) (i) (i) (i) (i) (i)
where P1 = (W1 C1 + V(i) )D1 + W1 C2 D2 and P2 which, by Lagrangian, becomes:
= W2(i) C(i) (i) (i)
1 D1 + (W2 C2 + V )D2 .
(i)

By MC simulation, the Jacobian recursions of (10)–(11) 1   

Hx − y2 + λ · TV(z) + μ z − x2
min 2 2
boil down to a simple noise evolution (13)–(14). We further x,z 2 2
have three remarks to facilitate the SURE-MC computations.
where μ is an augmented Lagrangian penalty parameter. The
– D1 n0 , D2 n0 , DT1 n0 and DT2 n0
can be computed by simple ADMM alternatively minimizes this functional w.r.t. both
first-order differences. variables x and z (iterate on i):
– The diagonal matrix–vector multiplications Ws(i) n0 , C(i) *
s
 2  2
n0 (s = 1, 2) and V(i) n0 are essentially simple compo- x(i) = arg minx 21 Hx − y2 + μ2 x − z(i) 2
 2
nentwise products: there is no need to explicitly write out z(i+1) = arg minz μ2 z − x(i) 2 + λ · TV(z)
the full diagonal matrices.
– For simplicity, we initialize u(0) = 0, and hence x(0) = i.e.
y. Thus, Jy (u(0) ) = 0 and Jy (x(0) ) = I, which yields
(0) (0) *  −1  T 
nu = 0 and nx = n0 . x(i) = HT H + μI H y + μz(i)
  (15)
z(i+1) = arg minz 21 z − x(i) 2 + μλ · TV(z)
2
Thus, all the computations of (13)–(14) can be efficiently
performed by element-wise operations (e.g. scalar difference,
multiplication, etc). We are able to evaluate the SURE with- We found that the update of z(i) is essentially a TV denois-
out any explicit matrix computation, summarized as follows. ing problem: estimate a ‘denoised’ version of a ‘noisy’ image
x(i) . It can be efficiently obtained by Chambolle’s algorithm
[2] (with the index j, for the fixed i)3 :
Algorithm 2: Monte Carlo counterpart of Algorithm 1
for i = 1, 2, ... (Chambolle’s iteration) do ⎛ ⎞
w(i, j)
1 compute x(i) by (6);   
⎜ τ μ  (i) λ T (i, j) ⎟
(i) (i) (i) (i, j) ⎜ (i, j) ⎟
2 compute nu1 , nu2 and nx by (13) and (14); u(i, j+1) =V ⎜u − D x + D u ⎟ (16)
3 compute the trace of Jy (x(i) ) by (12); ⎝ λ μ ⎠
4 compute SURE of ith iterate by (7);   
end z(i, j)

where diagonal block V(i, j) is given by:

To find the optimal value of λ, we repeatedly implement
 −1
Algorithm 2 for various tentative values of λ, and then, the (i, j) τμ  2  2
minimum SURE indicates the optimal λ. This global search Vn,n = 1 + (D1 z(i, j) )n + (D2 z(i, j) )n + α
λ
has been frequently used in [18,23].

When the Chambolle’s iteration (inner iterate on j)

3 Recursive evaluation of SURE for TV reaches the final convergence, z(i+1) is updated as z(i+1) =
deconvolution z(i,∞) .

Following the similar procedure with Sect. 2, we now con- 3.2 Jacobian recursion of ADMM
sider the TV deconvolution problem, where the SURE also
requires to compute the solution 
xλ and its SURE. For the deconvolution problem, the SURE needs to compute
H−1 in (5). However, it is observed that for the ill-conditioned
3.1 Basic scheme of ADMM matrix H, the simple inversion H−1 may cause numerical

To find 
xλ , we choose a typical alternating direction method 3We can see that the Chambolle’s iteration is readily incorporated into
of multipliers (ADMM) for TV deconvolution [i.e. H being ADMM, see Sect. 2.1 for details.

123
Signal, Image and Video Processing

instability of SURE [22]. Hence, we use the regularized input given y and H
inverse Hβ−1 to replace H−1 : initial z(0)

Chambolle
Hβ−1 = (HT H + βI)−1 HT update x(i) compute z(i)
by (15) by (16)
compute SURE
with a parameter β. The regularized SURE becomes: by (19), (17)
(i) (i)
update nx compute nz
1     by (20) by (20) ADMM
SURE = x(i) 2 − 2yT H−T x(i) + 2σ 2 Tr H−T Jy (x(i) )
N 2 β β ADMM iteration by i := i + 1
1 2
+ x0 2 (17)
N output xλ and SURE

Refer to [18,22] for the similar treatment. Fig. 1 SURE-MC evaluation for ADMM (Chambolle’s algorithm is
By the similar derivations with Sect. 2.2, we obtain the for obtaining z(i) )
Jacobian recursions for ADMM as:
⎧  
⎪ Jy (x(i) ) = (HT H + μI)−1 HT + μJy (z(i) ) 4 Experimental results and discussion
⎪
⎪
⎨ J (u(i, j+1) ) = V(i, j) J (u(i, j) ) − τ P(i, j) J (z(i, j) )
y 1 y 1 λ 1 y
(i, j+1) (i, j) (i, j) (18) 4.1 Experimental setting
⎪
⎪ Jy (u2 ) = V(i, j) Jy (u2 ) − τλ P2 Jy (z(i, j) )
⎪
⎩ J (z(i, j) ) = J (x(i) ) + λ DT J (u(i, j) )
y y μ y
The test dataset contains four 8-bit images of size 256 × 256
(i, j) (i, j) (i, j) (i, j) (i, j)
or 512 × 512 displayed in Fig. 2, covering a wide range of
where P1 = (W1 C1 + V(i, j) )D1 + W1 C2 D2 natural images.
(i, j) (i, j) (i, j) (i, j) (i, j)
and P2 = W2 C1 D1 + (W2 C2 + V(i, j) )D2 . For both denoising and deconvolution, we always termi-
(i, j) (i, j) (i, j) (i, j) nate the iterative algorithms, when the relative error of the
Here, W1 , W2 , C1 and C2 are defined similarly
with Sect. 2.2. objective value L(x(i) ) in (2) is below 10−5 .
The restoration performance is measured by the peak
3.3 Monte Carlo for SURE evaluation signal-to-noise ratio (PSNR), defined as (in dB) [20,22]:

 
Similar to Sect. 2.3, we adopt Monte Carlo to evaluate the 2552
PSNR = 10 × log10
trace term of SURE as: 
x − x0 22 /N
⎧ ⎫
⎪
⎪ ⎪
⎪
 −T  ⎨ ⎬ We choose α = 10−12 in the TV definition of (3) and set
(i) T −T (i)
Tr Hβ Jy (x ) = E n0 Hβ Jy (x )n0 (19) the parameter τ = 1/4 in (6), as suggested in [2].
⎪
⎪   ⎪ ⎪
⎩ (i)
⎭
nx
4.2 Image denoising
with input white Gaussian noise n0 ∼ N (0, I N ). Then,
multiplying n0 on both sides of (18), we obtain the noise 4.2.1 SURE evaluation for Chambolle’s algorithm
evolution during ADMM:
⎧ (i) (i)
First, we need to verify the accuracy of SURE w.r.t. MSE for
⎪
⎪ nx = B−1 HT n0 + μB−1 nz the Chambolle’s iteration. Figure 3 shows the convergence
⎪
⎨ n(i, j+1) (i, j) (i, j) (i, j)
u1 = V(i, j) nu1 − τλ P1 nz of Chambolle and evolution of SURE, under the noise levels
(i, j+1) (i, j) (i, j) (i, j) (20)
⎪
⎪ n = V(i, j) nu2 − τλ P2 nz of σ 2 = 1, 10 and 100, respectively. We can see that: (1)
⎪ u2
⎩ (i, j) (i, j) (i, (i, j)
nx + μλ DT1 nu1 + μλ DT2 nu2
j)
nz =

where B = HT H + μI, nx(i) = Jy (x(i) )n0 and other noises

are similarly defined as nx(i) .
The flowchart of the algorithm is shown in Fig. 1. Besides
from the three remarks mentioned in Sect. 2.3, we notice that
B−1 , HT and Hβ−1 can be computed by Fourier transform.
Thus, all computations of (20) can be performed without Fig. 2 Original test images: a Cameraman 256 × 256; b Coco 256 ×
any explicit matrix computations. 256; c House 256 × 256; d Bridge 512 × 512

123
 Signal, Image and Video Processing

(1) Cameraman, σ 2 = 1 (2) House, σ 2 = 10 (3) Bridge, σ 2 = 100 noisy images

objective value of L(x(i) ) — converegence of Chambolle’s iteration PSNR=28.13dB PSNR=28.13dB PSNR=18.13dB PSNR=18.13dB
96000 1 800000 1 8.00E+007 1
objective value objective value objective value
error of obj. value 0.1 error of obj. value error of obj. value
0.1 0.1

error of obj. value

error of obj. value

error of obj. value

objective value

objective value

objective value
7.00E+007
0.01 750000
95500 0.01 0.01

1E-3
6.00E+007
1E-3 1E-3
1E-4 700000
95000 1E-4 1E-4
1E-5 5.00E+007

1E-6 1E-5 1E-5

650000
1 2 3 4 5 4 8 12 16 20 15 30 45 60 75
iteration number iteration number iteration number

evolutions of SURE and MSE of x(i) during Chambolle’s iteration denoised images by SURE
1.00 MSE
10
MSE 100
PSNR=32.72dB PSNR=36.19dB PSNR=29.20dB PSNR=24.76dB
SURE SURE
0.98 9 90
MSE/SURE

MSE/SURE

MSE/SURE
0.96
8 80
0.94

0.92 7 70

0.90
60
MSE
6
SURE
0.88
1 2 3 4 5 4 8 12 16 20 15 30 45 60 75
iteration number iteration number iteration number

Fig. 3 The convergence of Chambolle and evolution of SURE for fixed

λ: (1) λ = 0.1; (2) λ = 1; (3) λ = 10 Fig. 6 Examples of visual comparisons: (1) Cameraman, σ 2 = 100;
(2) Coco, σ 2 = 100; (3) House, σ 2 = 1000; (4) Bridge, σ 2 = 1000
(1) Cameraman, σ 2 = 1 (2) House, σ 2 = 10 (3) Bridge, σ 2 = 100
2.0 100
MSE MSE
20
SURE SURE
1.8
We repeatedly implement the algorithm for various val-
MSE/SURE

MSE/SURE

MSE/SURE

90

1.6 16

1.4
opt. λ = 0.21 by SURE min.
opt. λ = 0.21 by MSE min. 12
opt. λ = 1.39 by SURE min.
opt. λ = 1.39 by MSE min.
80
opt. λ = 4.29 by SURE min.
opt. λ = 4.29 by MSE min. ues of λ and obtain Fig. 4, where the optimal λ is easy to
70
1.2

1.0 8
60 MSE
recognize.
SURE
0.8
1E-3 0.01 0.1 1 0.01 0.1 1 10 0.1 1 10
regularization parameter regularization parameter regularization parameter

4.2.2 Comparisons with discrepancy principle

Fig. 4 The global optimization of λ for TV denoising by Chambolle’s
algorithm
Discrepancy principle (DP) believes that a good value of λ
 2
(1) Cameraman, σ = 100 (2) House, σ = 10 (3) Bridge, σ = 1000
2 2 2 should satisfy the discrepancy condition y −xλ 2 = N σ 2 ,
objective value of L(x(i) ) — converegence of Chambolle’s iteration
objective value
1 1.8E6
objective value
1
objective value
1
according to the observation model (1) [8]. The original
9E6
error of obj. value error of obj. value 2.4E8
error of obj. value
Chambolle’s algorithm applied DP to update the parameter
error of obj. value

error of obj. value

error of obj. value

0.1 0.1 0.1

objective value

objective value

objective value

1.5E6

λ during iterations as [2]:

0.01 0.01 1.8E8 0.01
6E6 1.2E6
1E-3 1E-3 1.2E8 1E-3

.
9E5 1E-4
1E-4 1E-4
3E6 6E7

Nσ2
1E-5 1E-5 1E-5
6E5
50 100 150 50 100 150 200 250

(i+1)
λ(i)
50 100 150 200
iteration number iteration number

λ =
iteration number

update of λ during Chambolle’s iteration

10
2.8
30
y − x(i) 22
8
2.4
24
update of λ

update of λ

update of λ

6
2.0
18
which finally satisfies the discrepancy condition when con-
4
λ(i) → 8.37 1.6
λ(i) → 2.42 12 λ(i) → 29.55
2
1.2 6
verged. Figure 5 shows a few examples of the parameter
50 100
iteration number
150 200
0.8
50 100
iteration number
150 50 100 150
iteration number
200 250 update by DP.
Table 1 shows the complete comparisons between the pro-
Fig. 5 The convergence of Chambolle with the parameter update by posed SURE-based method and discrepancy principle. Here,
discrepancy principle
‘DP’ denotes discrepancy principle. The format of this table
est. λ , where the upper value is the selected value of λ by
is PSNR
the objective value keeps decreasing until convergence; (2) DP/SURE/MSE and the lower one is its resultant denoising
SURE is always close to MSE during the iterations. PSNR (in dB) using the corresponding λ. The symbol ‘–’

Table 1 The complete

σ2 1 10 100 1000 1 10 100 1000
comparisons of selected value of
λ and denoising PSNR between Image Cameraman House
DP and SURE
0.64 2.32 8.37 31.21 0.67 2.42 9.97
DP [2] 46.97 39.11 32.05 26.23 47.01 39.54 33.89 –
0.21 1.15 5.70 23.36 0.19 1.26 6.87 29.76
SURE 48.72 40.22 32.74 26.61 48.74 40.50 34.35 29.20
0.21 1.26 5.70 23.36 0.21 1.39 6.87 29.76
MSE 48.72 40.23 32.74 26.61 48.75 40.50 34.35 29.20

Image Coco Bridge

0.73 3.25 13.78 0.64 2.19 7.83 29.55
DP [2] 47.71 41.25 35.15 – 45.83 36.95 29.52 24.27
0.28 2.02 7.54 29.76 0.11 0.72 4.29 20.34
SURE 49.13 42.02 36.19 30.62 48.25 38.76 30.60 24.76
0.28 1.84 8.29 29.76 0.10 0.72 4.29 20.23
MSE 49.13 42.04 36.21 30.62 48.25 38.76 30.60 24.76

123
Signal, Image and Video Processing

(1) Cameraman
Rational, BSNR=40dB
(2) Coco
Uniform, BSNR=30dB
(3) Bridge
Gaussian, BSNR=10dB indicates that the method fails to find an optimal λ for this
objective value of L(x(i) ) — converegence of ADMM
120000
objective value
10
450000 objective value
1 9.00E+007
objective value 0.1
case. The MSE is not accessible in practice and thus shown
110000
error of obj. value 1
420000
error of obj. value
0.1
error of obj. value
in italics. It is the comparison benchmark, indicating the best

error of obj. value

error of obj. value

error of obj. value

0.01
objective value

objective value

objective value
0.1 7.50E+007
100000 390000 0.01

90000
0.01
360000 1E-3
1E-3 PSNR performance we can achieve. We can see that com-
1E-3 6.00E+007

80000
1E-4
330000
1E-4
1E-4
pared to the MSE minimization, the PSNR by DP is worse
300000
70000

5 10 15 20 25 30 35 40
1E-5
10 20 30 40 50
1E-5 4.50E+007

50 100 150
iteration number
200 250
1E-5
than optimal PSNR by 1 dB in average, whereas the SURE
iteration number iteration number

160
evolutions of SURE and MSE of x(i) during ADMM iteration minimization yields negligible PSNR loss (within 0.02 dB).
600
MSE MSE MSE

140
SURE
100 SURE
550
SURE
Figure 6 shows a number of visual examples.
MSE/SURE

MSE/SURE

MSE/SURE
120 80
500

100

80
60
450

400
4.3 Image deconvolution
40
60 350
5 10 15 20 25 30 35 40 10 20 30 40 50 50 100 150 200 250
iteration number iteration number iteration number
4.3.1 Experimental setting
Fig. 7 The convergence of ADMM and evolution of SURE for fixed λ:
(1) λ = 0.1, (2) λ = 1, (3) λ = 10 For deconvolution problem, we consider the following
benchmark convolution kernels commonly used in [10,22]:
(1) Cameraman (2) Coco (3) Bridge
Rational, BSNR=40dB Uniform, BSNR=30dB Gaussian, BSNR=10dB

– Rational filter h(i, j) = C · (1 + i 2 + j 2 )−1 for

250
MSE 140 MSE MSE
700
SURE SURE SURE
200 120

i, j = −7, . . . , 7;
MSE/SURE

MSE/SURE

MSE/SURE

600

150 λ = 0.0168 by SURE min. 100

λ = 0.0672 by λ = 2.04 by
λ = 0.0146 by MSE min. 80 SURE and MSE min. 500 SURE and MSE min.
100
60
400
– Separable filter 5 × 5 filter with weights [1, 4, 6, 4, 1]
/16 along both horizontal and vertical directions;
50 40

300
20
1E-3 0.01 0.1 1 0.01 0.1 1 0.1 1 10
regularization parameter regularization parameter regularization parameter
– 9 × 9 uniform blur;
 2+ j 2 
Fig. 8 The global optimization of λ for TV deconvolution by ADMM – Gaussian kernel h(i, j) = C · exp − i 2s 2 with
s = 2.0.
(1) Cameraman (2) Coco (3) Bridge
Separable, BSNR=30dB Uniform, BSNR=20dB Gaussian, BSNR=10dB
/
where C is a normalization factor, s.t. i, j h(i, j) = 1.0.
1 0.14

4.5E5 3.28E7
9.0E5
objective value

objective value

objective value

λ(i) → 1.66 λ(i) → 0.13

1
update of λ

update of λ

0.12

The blurred images are subsequently contaminated by i.i.d

update of λ

3.24E7

λ(i) → 0.51
3.0E5 0.1 7.5E5

6.0E5
0.1
3.20E7 0.1
Gaussian noise with various variance σ 2 , corresponding to
3.16E7
1.5E5

10 20 30
objective value
update of λ
40 50
0.01
20 40 60
objective value
update of λ
80 100
0.01
5 10 15 20
objective value
update of λ
25 30
0.08 blur signal-to-noise ratio (BSNR) being 40, 30, 20 and 10 dB,
iteration number iteration number iteration number

respectively, where the BSNR is defined as (in dB) [20]:

Fig. 9 The parameter update for TV deconvolution by DP [10]

Table 2 The selected λ and

Blur kernel Rational filtering Separable filtering
corresponding PSNR by DP and
SURE (rational and separable BSNR (in dB) 40 30 20 10 40 30 20 10
filtering)
Cameraman
0.05 0.20 0.54 0.18 0.12 0.51 4.90 2.15
DP [10] 31.35 27.84 24.64 19.81 30.55 28.81 26.85 23.69
0.01 0.05 0.28 3.59 0.01 0.10 0.60 3.60
SURE 32.75 28.26 24.99 22.58 31.44 29.44 27.10 24.31
0.02 0.06 0.36 4.64 0.02 0.10 0.77 3.60
MSE 32.81 28.31 25.05 22.59 31.49 29.44 27.13 24.31
Coco
0.21 0.48 1.64 4.83 0.96 1.42 3.70 6.82
DP [10] 38.59 35.76 31.62 28.80 40.37 38.03 33.36 31.78
0.04 0.08 0.77 5.99 0.02 0.10 0.61 7.74
SURE 39.92 36.13 32.50 28.73 41.30 38.58 34.98 32.06
0.03 0.11 0.60 3.59 0.04 0.12 0.72 7.74
MSE 39.95 36.35 32.55 28.89 41.35 38.70 35.23 32.06
House
0.08 0.29 1.03 2.77 0.36 0.75 2.57 4.74
DP [10] 35.83 33.39 29.81 27.03 36.33 34.47 32.43 29.01
0.02 0.11 0.60 3.59 0.01 0.07 1.29 12.92
SURE 37.05 33.79 30.64 27.00 37.55 35.42 32.91 29.80
0.02 0.08 0.60 2.78 0.01 0.10 1.29 11.57
MSE 37.14 33.81 30.64 27.03 37.56 35.55 32.91 30.00
Bridge
0.03 0.15 0.34 0.18 0.06 0.34 1.09 9.24
DP [10] 30.00 26.97 24.75 20.61 29.42 28.00 26.38 24.33
0.01 0.05 0.17 1.67 0.007 0.06 0.36 3.59
SURE 31.05 27.57 24.90 22.54 30.16 28.58 26.71 24.49
0.008 0.05 0.22 2.78 0.007 0.06 0.46 3.59
MSE 31.06 27.57 24.97 22.72 30.16 28.58 26.74 24.49

123
 Signal, Image and Video Processing

Table 3 The selected λ and

Blur kernel 9 × 9 uniform Gaussian kernel
corresponding PSNR by DP and
SURE (uniform and Gaussian BSNR (in dB) 40 30 20 10 40 30 20 10
blurs)
Cameraman
0.03 0.002 0.007 0.01 0.05 0.007 0.02 0.12
DP [10] 28.04 24.44 22.46 19.22 26.06 24.67 23.32 19.89
0.01 0.08 0.11 7.74 0.01 0.11 0.28 2.78
SURE 28.42 25.44 22.98 21.29 26.08 25.02 23.86 22.57
0.01 0.08 0.17 2.78 0.01 0.08 0.46 4.64
MSE 28.43 25.44 23.02 21.43 26.09 25.06 23.87 22.67
Coco
0.10 0.34 1.66 0.05 0.12 0.47 0.07 0.14
DP [10] 35.47 32.94 28.35 23.62 35.26 33.01 29.96 23.94
0.02 0.08 0.60 2.78 0.01 0.06 0.60 5.99
SURE 36.54 33.54 30.02 27.46 35.36 33.60 31.67 29.29
0.01 0.10 0.60 2.15 0.03 0.10 0.77 4.64
MSE 36.61 33.59 30.02 27.46 35.63 33.74 31.68 29.34
House
0.04 0.006 0.02 0.02 0.08 0.34 0.03 0.12
DP [10] 34.66 29.52 26.74 22.75 32.44 30.54 28.23 23.45
0.008 0.08 0.77 2.15 0.01 0.08 0.77 2.15
SURE 35.04 32.05 28.41 25.41 32.59 31.27 29.62 27.30
0.01 0.08 0.46 1.67 0.03 0.10 0.77 2.15
MSE 35.10 32.05 28.52 25.45 32.82 31.34 29.62 27.42
Bridge
0.02 0.32 0.007 0.01 0.03 0.008 0.02 0.13
DP [10] 26.80 25.07 22.99 20.01 25.66 24.85 23.70 19.36
0.006 0.03 0.17 1.29 0.002 0.02 0.17 1.67
SURE 27.25 25.25 23.40 21.76 25.99 25.09 24.03 22.70
0.005 0.03 0.13 1.67 0.002 0.01 0.17 2.78
MSE 27.26 25.25 23.42 21.80 25.99 25.10 24.03 22.81

observed images
PSNR=22.25dB PSNR=26.33dB PSNR=29.59dB PSNR=20.49dB 4.3.2 SURE evaluation for ADMM

First, we implement the procedure shown in Fig. 1, i.e. apply

ADMM to solve (2) with fixed λ, and evaluate the SURE.
Figure 7 shows the convergence of ADMM and the evolution
of SURE. Figure 8 shows the global optimization of λ.
deconvolved images by DP
PSNR=31.35dB PSNR=32.94dB PSNR=32.43dB PSNR=19.36dB

4.3.3 Comparisons with discrepancy principle

Similar to [2], the DP has also been adopted to the deconvo-

lution problem, for example, the i-LET method updated the
deconvolved images by SURE parameter λ by the following rule:
PSNR=32.75dB PSNR=33.54dB PSNR=32.91dB PSNR=22.70dB

Nσ2
λ(i+1) = λ(i)
y − Hx(i) 22

in 1 -based sparse deconvolution [10]. We use this equation

Fig. 10 Examples of visual comparisons: (1) Cameraman, rational, to update λ in TV deconvolution and obtain Fig. 9.
BSNR = 40 dB; (2) Coco, uniform, BSNR = 30 dB; (3) House, sepa- Tables 2 and 3 show the comparisons between the pro-
rable, BSNR = 20 dB; (4) Bridge, Gaussian, BSNR = 10 dB
posed SURE and DP. The format is the same as Table 1. We
can see that compared to the best PSNR results obtained by
0  1 MSE minimization, the PSNR by DP is worse than optimal
Hx0 − mean(Hx0 )2
BSNR = 10 × log10 2 PSNR by 1–3 dB in average, whereas the SURE minimiza-
Nσ2 tion yields negligible PSNR loss (at most 0.20 dB). Figure 10
shows a number of visual examples, where we can see the
The deconvolution performance is also measured by PSNR. better visual quality by SURE than by DP.4
In addition, we always choose the parameters μ = 0.1σ 2
and β = 10−5 .
4 It is better to recognize the visual difference by zoom-in on larger
screen.

123
Signal, Image and Video Processing

5 Conclusions 11. Ramirez, C., Argaez, M.: An 1 minimization algorithm for non-
smooth regularization in image processing. Signal Image Video
Process. 9, 373–386 (2015)
In this paper, we presented a SURE-based method for 12. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based
automatically tuning regularization parameter for TV-based noise removal algorithms. Physica D 60, 259–268 (1992)
recovery. In particular, we proposed a recursive evaluation 13. Selesnick, I., Parekh, A., Bayram, I.: Convex 1-D total variation
and Monte Carlo simulation for the practical computation. denoising with non-convex regularization. IEEE Signal Process.
Lett. 22(2), 141–144 (2015)
Numerical results showed the superior performance of SURE 14. Shen, C., Bao, X., Tan, J., Liu, S., Liu, Z.: Two noise-robust axial
to other criteria for parameter selection, e.g. discrepancy scanning multi-image phase retrieval algorithms based on Pauta
principle. criterion and smoothness constraint. Opt. Express 25(14), 16235–
This proposed method, in principle, can be extended 16249 (2017)
15. Siadat, M., Aghazadeh, N., Öktem, O.: Reordering for improving
to more complicated (possibly non-convex) regularizers
global Arnoldi–Tikhonov method in image restoration problems.
[13,18,23]. Future work will also deal with the SURE- Signal Image Video Process. 12, 497–504 (2018)
based multiple parameter selection and faster optimization 16. Stein, C.M.: Estimation of the mean of a multivariate normal dis-
of SURE, to accelerate the global search used here. tribution. In: Stein, C.M. (ed.) The Annals of Statistics, vol. 9, No.
6, pp. 1135–1151 (1981)
17. Tao, M., Yang, J.: Alternating Direction Algorithms for Total Varia-
tion Deconvolution in Image Reconstruction. Optimization Online,
TR0918, Department of Mathmatics, Nanjing University (2009)
18. Vonesch, C., Ramani, S., Unser, M.: Recursive risk estimation for
References non-linear image deconvolution with a wavelet-domain sparsity
constraint. In: IEEE International conference on Image processing,
1. Blu, T., Luisier, F.: The SURE-LET approach to image denoising. pp. 665–668 (2008)
IEEE Trans. Image Process. 16(11), 2778–2786 (2007) 19. Wang, Y., Yang, J., Yin, W., Zhang, Y.: A new alternating mini-
2. Chambolle, A.: An algorithm for total variation minimization and mization algorithm for total variation image reconstruction. SIAM
applications. J. Math. Imaging Vis. 20, 89–97 (2004) J. Imaging Sci. 1(3), 248–272 (2008)
3. Chen, D.: Inext alternating direction method based on Newton 20. Xue, F., Blu, T.: A novel SURE-based criterion for parametric PSF
descent algorithm with application to Poisson image deblurring. estimation. IEEE Trans. Image Process. 24(2), 595–607 (2015)
Signal Image Video Process. 11, 89–96 (2017) 21. Xue, F., Du, R., Liu, J.: A recursive predictive risk estimate for
4. Dell’Acqua, P.: ν acceleration of statistical iterative methods for proximal algorithms. In: Proceedings of the 41st IEEE International
image restoration. Signal Image Video Process. 10, 927–934 (2016) Conference on Acoustics, Speech and Signal Processing, Shanghai,
5. Golub, G., Heath, M., Wahba, G.: Generalized cross-validation as a China, March 20–25, pp. 4498–4502 (2016)
method for choosing a good ridge parameter. Technometrics 21(2), 22. Xue, F., Luisier, F., Blu, T.: Multi-Wiener SURE-LET deconvolu-
215–223 (1979) tion. IEEE Trans. Image Process. 22(5), 1954–1968 (2013)
6. Guo, C., Li, Q., Wei, C., Tan, J., Liu, S., Liu, Z.: Axial multi-image 23. Xue, F., Yagola, A.G., Liu, J., Meng, G.: Recursive SURE for iter-
phase retrieval under tilt illumination. Sci. Rep. 7, 7562 (2017) ative reweighted least square algorithms. Inverse Probl. Sci. Eng.
7. Hansen, P.C.: Analysis of discrete ill-posed problems by means of 24(4), 625–646 (2016)
the L-curve. SIAM Rev. 34(4), 561–580 (1992)
8. Morozov, V.: Methods for Solving Incorrectly Posed Problems.
Springer, New York (1984)
Publisher’s Note Springer Nature remains neutral with regard to juris-
9. Osher, S., Burger, M., Goldfarb, D., Xu, J., Yin, W.: An iterative
dictional claims in published maps and institutional affiliations.
regularization method for total variation-based image restoration.
SIAM J. Multiscale Model. Simul. 4(2), 460–489 (2005)
10. Pan, H., Blu, T.: An iterative linear expansion of thresholds for 1 -
based image restoration. IEEE Trans. Image Process. 22(9), 3715–
3728 (2013)

123