An Alternating Direction Method for Mixed Gaussian Plus Impulse Noise Removal

Abstract

A combined total variation and high-order total variation model is proposed to restore blurred images corrupted by impulse noise or mixed Gaussian plus impulse noise. We attack the proposed scheme with an alternating direction method of multipliers (ADMM). Numerical experiments demonstrate the efficiency of the proposed method and the performance of the proposed method is competitive with the existing state-of-the-art methods.

1. Introduction

Image restoration is one of the most fundamental tasks in image processing and plays an important role in various areas of applied sciences [1]. In the literature of image restoration, there exist a lot of good methods dealing with images which are contaminated by single kind of noise, such as additive Gaussian noise or impulse noise. However, in many practical situations, the observed images are usually corrupted by mixed noise, for example, Gaussian plus impulse noise and Gaussian plus Poisson noise. In this work, we focus on Gaussian plus impulse noise. This kind of mixed noise is commonly caused by malfunctioning arrays in camera sensors or transmission errors [2]. We aim to find the unknown true image 𝑓∈𝑅^𝑛2 from the observed image 𝑔∈𝑅^𝑛2 defined by 𝑔=𝑁_imp𝐻𝑓+𝑛, (1) where 𝑁_imp denote the process of image degradation with impulse noise, 𝐻∈𝑅^𝑛2×𝑛2 is a blurring matrix which is assumed to be known, and 𝑛∈𝑅^𝑛2 is an additive zero-mean Gaussian white noise of variance 𝜎². For simplicity and without loss of generality, we assume that the underlying images have square domains. Based on the noise values, impulse noise can be classified as salt-and-pepper noise and random-valued noise. Suppose that [𝑑_min,𝑑_max] is the dynamic range of 𝑓 and 𝑓_𝑗,𝑘 ((𝑗,𝑘)∈Ω={1,2,…,𝑛}×{1,2,…,𝑛}) is the gray value of an image 𝑓 at location (𝑗,𝑘) [3]. The operator 𝑔=𝑁_imp(𝑓) is defined as follows.

(i) Salt-and-pepper noise: the gray level of 𝑔 at pixel location (𝑗,𝑘) is 𝑔_𝑗,𝑘=𝑑_min𝑠,withprobability2,𝑑_max𝑠,withprobability2,𝑓_𝑗,𝑘,withprobability1−𝑠, (2)  where 𝑠 is the noise ratio which defines the level of the salt-and-pepper noise.

(ii) Random-valued noise: the gray level of 𝑔 at pixel location (𝑗,𝑘) is 𝑔_𝑗,𝑘=𝑑_𝑗,𝑘𝑓,withprobability𝑟,_𝑗,𝑘,withprobability1−𝑟, (3)  where 𝑑_𝑗,𝑘 are uniformly distributed random numbers in [𝑑_min,𝑑_max] and 𝑟 is the noise ratio which defines the level of the random-valued noise.

It is obvious that it is more difficult to remove the random-valued impulse noise than salt-and-pepper noise since the random-valued impulse noise can be arbitrary number in [𝑑_min,𝑑_max]. As is well known, recovering 𝑓 form 𝑔 is an ill-conditioned problem. A general approach to compute a meaningful approximation is the regularization method. Recently, variational models have attracted a lot of interest in cleaning mixed noise [3–9]. The key idea of these methods is based on a two-approach: detect the location of the impulse corrupted pixels then proceed with the filtering phase. For example, Cai et al. [4] proposed a two-phase approach for (1). In the first phase, they identify the location of the impulse noise and remove them from the data set through median filter. In the second phase, they minimize a functional of the form _{(𝑗,𝑘)∈𝒰}𝐻𝑓−𝑔_𝑗,𝑘^𝑝+𝛽Φ(𝑥),𝑝=1,2, (4) where 𝒰 is the set of data samples that are likely to be uncorrupted with impulse noise and Φ(𝑥) is the Mumford-Shah regularization term [10]: Φ(𝑥)=_Ω⧵Γ∇𝑓²+𝛼𝛽_Γ𝑑𝜎, (5) where Γ is the edge set. Due to nonconvexity of Mumford-Shah regularizer, there exist numerous local minimums of (4). It is difficult to handle the Mumford-Shah functional. To overcome this difficulty, the Γ-convergence functional for Φ is used in [4]. The reconstruction performance is competitive; however the choice of 𝑝=1 or 𝑝=2 in (4) is depending on noise level and the computational performance is poor. In [5], Cai et al. accelerate the method in [4] and the computational time is much less than that in [4], although it is still highly depending on noise level. Recently, Li et al. [9] considered a functional with a content-dependent fidelity term and then proposed an iterative framelet-based approximation deblurring algorithm IFASDA. The advantage of IFASDA is that this algorithm is parameter free, which makes the proposed method more practical. We note that the total variation (TV) is not used in [4, 9]; however, the TV regularization is very popular [11] as its superiority of preserving sharp edges or object boundaries in the recovered images. The modified total variation minimization scheme is proposed in [3] to restore the images and an alternating minimization algorithm is employed to solve the proposed problem. In [8], a cost functional consisting of TV regularization term and 𝑙₂ and 𝑙₁ data fidelity terms is proposed to remove mixed Gaussian plus impulsive noise. Their numerical experiments showed that their method was very efficient and the reconstruction performance was quite competitive. However, as we know the TV norm transforms the smooth area to piecewise constants, the so-called staircase effect. In order to overcome this spurious effect, many high-order PDEs have been proposed [12–14] to solve this problem. Motivated by this, we propose a new model combining total variation and high-order total variation to restore the blurred images corrupted by Gaussian plus impulse noise. Our numerical experiments show the effectiveness of the proposed approach.

The outline of this paper is as follows. In Section 2, we briefly introduce the work of Huang et al. in [3] and the alternating direction method of multipliers (ADMM). In Section 3, we present the proposed algorithm. To demonstrate the effectiveness of the proposed method, we will show some numerical results on several test images in Section 4. Finally, some conclusion remarks are drawn in Section 5.

2. Brief Review of Related Methods

2.1. Review of Current Methods

In [3], Huang et al. proposed the following modified total variation minimization scheme to restore the blurred images corrupted by Gaussian plus impulse noise: min_{𝑓,𝑢(𝑗,𝑘)∈𝒰}𝐻𝑓−𝑔_𝑗,𝑘^𝑝+𝛼₁𝑓−𝑢²₂+𝛼₂‖𝑢‖_TV, (6) where 𝒰 is the set of data samples that are likely to be uncorrupted by impulse noise, 𝑝=1 (or 2), 𝛼₁ and 𝛼₂ are two positive regularization parameters, and ‖⋅‖_TV is the discrete total variation (TV) regularization term. The discrete TV of 𝑓 is defined by ‖𝑓‖_TV=∑_{1≤𝑗,𝑘≤𝑛}|(∇𝑓)^𝑥_𝑗,𝑘|²+|(∇𝑓)^𝑦_𝑗,𝑘|². The discrete gradient operator ∇∶𝑅^𝑛2→𝑅^𝑛2×2 is defined by (∇𝑓)_𝑗,𝑘=((∇𝑓)^𝑥_𝑗,𝑘,(∇𝑓)^𝑦_𝑗,𝑘) with ∇𝑓^𝑥_𝑗,𝑘=𝑓_𝑗+1,𝑘−𝑓_𝑗,𝑘if𝑗<𝑛,0if𝑗=𝑛,∇𝑓^𝑦_𝑗,𝑘=𝑓_𝑗,𝑘+1−𝑓_𝑗,𝑘if𝑘<𝑛,0if𝑘=𝑛, (7) for 𝑗,𝑘=1,…,𝑛. Here 𝑓_𝑗,𝑘 refers to the ((𝑘−1)𝑛+𝑗)th entry of the vector 𝑓 (it is the (𝑗,𝑘)th pixel location of the image). In [3], the median-type filter was used to identify the location of the possible noisy pixels in the first phase; then problem (6) was solved by alternating minimization method with respect to 𝑓 and 𝑢. As reported in [3], they used the 𝐿₁ norm in the experiment, and the restoration results by using 𝐿₂ norm were about the same as those by using the 𝐿₁ norm for impulse plus Gaussian noise condition.

2.2. The Alternating Direction Method of Multipliers (ADMM)

In this section, we present the alternating direction method of multipliers (ADMM) [15–17] which belongs to the family of augmented Lagrangian methods. Consider the following linearly constrained separable convex minimization problem of the form: min𝜃₁𝑥₁+𝜃₂𝑥₂,𝐴₁𝑥₁+𝐴₂𝑥₂𝑥=𝑏,₁∈Ω₁,𝑥₂∈Ω₂, (8) where 𝜃₁∶ℝ^𝑛1→ℝ and 𝜃₂∶ℝ^𝑛2→ℝ are closed proper convex functions, Ω₁⊆ℝ^𝑛1 and Ω₂⊆ℝ^𝑛2 are closed convex sets, 𝐴₁∈ℝ^𝑙×𝑛1 and 𝐴₂∈ℝ^𝑙×𝑛2 are given matrices, and 𝑏∈ℝ^𝑙 is a given vector. Specifically, the iterative scheme of ADMM for solving (8) is as follows.

Algorithm ADMM(1)Set  𝑘=0 and choose 𝛽>0, 𝑥⁰₂, and 𝑑⁰(2)repeat (3)𝑥₁^𝑘+1=argmin{𝜃₁(𝑥₁)+(𝛽/2)‖(𝐴₁𝑥₁+𝐴₂𝑥^𝑘₂−𝑏)−𝑑^𝑘‖²∣𝑥₁∈Ω₁},(4)𝑥₂^𝑘+1=argmin{𝜃₂(𝑥₂)+(𝛽/2)‖(𝐴₁𝑥₁^𝑘+1+𝐴₂𝑥₂−𝑏)−𝑑^𝑘‖²∣𝑥₂∈Ω₂}, (5)𝑑^𝑘+1=𝑑^𝑘−(𝐴₁𝑥₁^𝑘+1+𝐴₂𝑥₂^𝑘+1−𝑏),(6)𝑘←𝑘+1,(7)until stopping criterion is satisfied.

As reported in [18], the penalty parameter 𝛽 can be substituted by a symmetric positive definite (spd) matrix 𝑄 and the extension of ADMM is still convergent. This leads to the following extension of algorithm ADMM.

Extension of Algorithm ADMM(1)Set  𝑘=0 and choose 𝛽>0, 𝑥⁰₂, and 𝑑⁰, (2)repeat (3)𝑥₁^𝑘+1=argmin{𝜃₁(𝑥₁)+(1/2)‖(𝐴₁𝑥₁+𝐴₂𝑥^𝑘₂−𝑏)−𝑑^𝑘‖²_𝑄∣𝑥₁∈Ω₁},(4)𝑥₂^𝑘+1=argmin{𝜃₂(𝑥₂)+(1/2)‖(𝐴₁𝑥₁^𝑘+1+𝐴₂𝑥₂−𝑏)−𝑑^𝑘‖²_𝑄∣𝑥₂∈Ω₂},(5)𝑑^𝑘+1=𝑑^𝑘−(𝐴₁𝑥₁^𝑘+1+𝐴₂𝑥₂^𝑘+1−𝑏),(6)𝑘←𝑘+1, (7)until stopping criterion is satisfied.

3. The Proposed Approach

Throughout this paper, we will assume that the set of outlier (pixels corrupted with impulse noise) is known. Actually, we can use the adaptive median filter (AMF) [19] to detect salt-and-pepper noise and the adaptive center-weighted median filter (ACWMF) [20] for random-valued impulse noise case. Using the set 𝒰 represents the data samples that are likely to be uncorrupted by impulse noise; then we proposed the following objective function; min_𝑓Λ𝐻𝑓−𝑔^𝑝_𝑝+𝛼₁𝑓_TV+𝛼₂𝑓_HTV, (9) where Λ=diag(𝐈_{[𝑓∈𝒰]}), 𝐈_[⋅] is the indicator function, 𝑝=1 (or 2), and ‖𝑓‖_HTV=∫_Ω(𝑓²_𝑥𝑥+𝑓²_𝑥𝑦+𝑓²_𝑦𝑥+𝑓²_𝑦𝑦)^1/2𝑑𝑥𝑑𝑦 [12]. For simplicity we introduce the notation |𝐷²𝑓|=(𝑓²_𝑥𝑥+𝑓²_𝑥𝑦+𝑓²_𝑦𝑥+𝑓²_𝑦𝑦)^1/2. Then we will discuss the details of the algorithm to solve the model (9). For convenience we introduce three new variables and rewrite (9) in the constrained optimization problem as follows: min_{𝑣1,𝑣2,𝑣3,𝑓,𝑑}Λ𝑣₁^𝑝_𝑝+𝛼₁𝑣_2TV+𝛼₂𝑣₃₁∣𝑣₁=𝐻𝑓−𝑔,𝑣₂=𝑓,𝑣₃=𝐷²𝑓. (10)

We apply the extension of ADMM algorithm to the constrained problem (10), which we call mixed noise image deblurring by extension of ADMM (MNID-ADMM). In the following algorithm, we can simply choose positive matrix 𝑄 as diag(𝛽₁𝐼,𝛽₂𝐼,𝛽₃𝐼), where 𝐼 is the identity matrix. If we set 𝑄=𝛽𝐼, it is known as ADMM algorithm.

Algorithm of MNID-ADMM(1)Set  𝑘=0, 𝑑⁰₁,𝑑⁰₂,𝑑⁰₃=0, and choose 𝛼₁>0, 𝛼₂>0 and positive matrix 𝐻, (2)repeat(3)𝑣₁^𝑘+1=argmin_𝑣1‖Λ𝑣₁‖^𝑝_𝑝+(𝛽₁/2)‖𝑣₁−(𝐻𝑓^𝑘−𝑔)−𝑑^𝑘₁‖²,(4)𝑣₂^𝑘+1=argmin_𝑣2(𝛽₂/2)‖𝑣₂−𝑓^𝑘−𝑑^𝑘₂‖²₂+𝛼₁‖𝑣₂‖_TV, (5)𝑣₃^𝑘+1=argmin_𝑣3𝛼₂‖𝑣₃‖₁+(𝛽₃/2)‖𝑣₃−𝐷²𝑓^𝑘−𝑑^𝑘₃‖²₂,(6)𝑓^𝑘+1=argmin_𝑓(𝛽₁/2)‖𝐻𝑓−𝑔+𝑑^𝑘₁−𝑣₁^𝑘+1‖²₂+(𝛽₂/2)‖𝑓+𝑑^𝑘₂−𝑣₂^𝑘+1‖²₂+(𝛽₃/2)‖𝐷²𝑓+𝑑^𝑘₃−𝑣₃^𝑘+1‖²₂, (7)𝑑₁^𝑘+1←𝑑^𝑘₁−(𝑣₁^𝑘+1−(𝐻𝑓^𝑘+1−𝑔)), 𝑑₂^𝑘+1←𝑑^𝑘₂−(𝑣₂^𝑘+1−𝑓^𝑘+1),  𝑑₃^𝑘+1←𝑑^𝑘₃−(𝑣₃^𝑘+1−𝐷²𝑓),(8)𝑘←𝑘+1,(9)until stopping criterion is satisfied.

Cai et al. [4] considered to use 𝑙_𝑝 (𝑝=1 or 𝑝=2) as the data-fidelity term and discussed how to choose 𝑙₁ or 𝑙₂ norm. It is known that 𝑙₁ data-fidelity term is more suitable than 𝑙₂ data-fidelity when applying to remove impulse noise [21, 22]. From an experimental point of view, we find that it is faster to use 𝑙₂ norm than 𝑙₁ norm. Considering this, we use 𝑙₂ data-fidelity term when the images are corrupted by mixed Gaussian plus impulse noise and we use 𝑙₁ data-fidelity term when the images are corrupted by impulse noise only. Then, we will discuss the details of solving the 𝑣₁, 𝑣₂, 𝑣₃, and 𝑓 subproblems, respectively. For 𝑣₁ subproblem, when 𝑝=1, it is easy to solve the 𝑣₁ subproblem through shrinkage and the minimizer 𝑣₁ is given by 𝑣₁^𝑘+1_𝑖=𝐻𝑓^𝑘−𝑔+𝑑^𝑘_1𝑖ifΛ(𝑖,𝑖)=0,max𝐻𝑓^𝑘−𝑔+𝑑^𝑘_1𝑖−1𝛽₁,0×sgn𝐻𝑓^𝑘−𝑔+𝑑^𝑘_1𝑖ifΛ(𝑖,𝑖)=1. (11) When 𝑝=2, the minimizer 𝑣₁ is given by 𝑣₁^𝑘+1=(2Λ+𝛽₁𝐼)⁻¹𝛽₁(𝐻𝑓^𝑘−𝑔+𝑑^𝑘₁). There are numerous algorithms to solve the 𝑣₂ subproblem [11, 23–25]; here, we consider adopting Chambolle’s algorithm [25]. The minimizer 𝑣₃ is given by 𝑣₃^𝑘+1=shrink_𝛼2/𝛽3𝑑^𝑘₃+𝐷²𝑓^𝑘, (12) where shrink_𝛾(𝑝)=max{0,1−(𝛾/‖𝑝‖₁)}𝑝. In order to get 𝑓^𝑘+1, we need to solve the following problem: 𝛽₁𝐻^𝑇𝐻+𝛽₂𝐼+𝛽₃𝐷^2𝑇𝐷²𝑓=𝛽₁𝐻^𝑇𝑔+𝑣₁^𝑘+1−𝑑^𝑘₁+𝛽₂𝑣₂^𝑘+1−𝑑^𝑘₂+𝛽₃𝐷^2𝑇𝑣₃^𝑘+1−𝑑^𝑘₃. (13)

In image restoration, 𝐻 is usually a matrix of special structure. For example, 𝐻 is a block-circulant with circulant block (BCCB) matrix when periodic boundary conditions are applied to the image boundary. The matrix 𝐻 can be diagonalized by the discrete fast Fourier transform (FFT) matrix. There are some fast algorithms to solve the problem (13) for different boundary conditions. In this paper, we use periodic boundary conditions. Then the solution of subproblem (13) can be exactly and efficiently calculated via fast Fourier transform.

4. Numerical Results

In this section, numerical experiments are presented to demonstrate the performance of our proposed MNID-ADMM for test images “Lena,” “Einstein,” “Cameraman,” and “Boat.” All test images are 256×256 gray level images. The results are compared with those obtained by the two-phase method denoted by “Cai-TP” [4] and the fast restoration method proposed in [3]. The splitting-and-penalty method is employed to solve the modified total variation scheme in [3]. As we know, the ADMM approach is more efficient and stable than the splitting-and-penalty method [26, 27]. For a fair comparison, we solve the total variation scheme (just set 𝛼₂=0 in (9)) by ADMM rather than the splitting-and-penalty method used in [3]. For simplicity, we call this method “HADMM” hereafter. In order to measure the quality of the restored images by different methods, we introduce the mean squared error (MSE), and the peak signal-to-noise ratio (PSNR): 1MSE=𝑛^2𝑛2_𝑗=1𝑓𝑗−𝑔𝑗²,PSNR=20log₁₀MAX_𝑓√,MSE (14) where 𝑓, 𝑔, 𝑛² and MAX_𝑓 are the original image, the blurred image, the number of pixels of image, and the maximum possible pixel value of the image, respectively. The higher PSNR, the better quality of the restoration. All the experiments are performed using MATLAB 7.10.0 on a computer equipped with an Intel(R) Pentium(R) 2.80 GHz processor, with 4.00 GB of RAM, and running Windows 7. The stop criterion of all methods is set to be ‖𝑓^𝑘+1−𝑓^𝑘‖²₂/‖𝑓^𝑘‖²₂<5×10⁻⁴. In order to reduce the computational time in the search for good regularization parameters, we set 𝛽₁=𝛽₂ for impulse noise only situation. It is reasonable to set 𝛼₁+𝛼₂=1; for further details, see [12]. Moreover, we find that we can get good restoration result by just setting 𝛽₁=𝛽₂=𝛽₃ for impulse plus Gaussian noise situation. An out-of-focus kernel with radius 3 is used to blur the original images in the following examples. The parameters were hand tuned to give the best PSNR improvement. Similarly as in [4], we also use the same parameters for different images for the same convolution kernel and noise level.

Example 1. We restore images contaminated by impulse noise only. In this example, we consider the images corrupted by salt-and-pepper noise with level of 70% and random-valued noise with level of 55% separately. The four original images and the blurred images are shown in Figure 1. The comparisons of MNID-ADMM, Cai-TP, and HADMM methods are shown in Figures 2 and 3 and Table 1. From the restoration results, we get that the MNID-ADMM method performs better than Cai-TP both in restoration quality and computation times and we get a little higher PSNR than HADMM method for various images.

Table 1 

The PSNR (dB) and computing time (seconds) for the restoration results by MNID-ADMM, Cai-TP, and HADMM for blurred images contaminated by impulse noise only. The blurring kernel is the out-of-focus kernel of radius 3.

Figure 1 

Top: the original images. Middle: Blurred and noisy images by using an out-of-focus kernel with radius 3 and corrupted by salt-and-pepper noise with noise level 𝑠=70%. Bottom: blurred and noisy images by using an out-of-focus kernel with radius 3 and corrupted by random-valued noise with noise level 𝑟=55%.

Figure 2 

Images blurred with out-of-focus kernel of radius 3 and then corrupted by salt-and-pepper noise with noise level 𝑠=70%. Top: the restored images by MNID-ADMM and the parameters we used are [𝛼1=0.01, 𝛼2=0.005, 𝛽1=𝛽2=0.01, 𝛽3=5𝑒−6]. Middle: the restored images by Cai-TP. Bottom: the restored images by HADMM.

Figure 3 

Images blurred with out-of-focus kernel of radius 3 and then corrupted by random-valued noise with noise level 𝑟=55%. Top: the restored images by MNID-ADMM and the parameters we used are [𝛼1=0.002, 𝛼2=0.005, 𝛽1=𝛽2=0.02, 𝛽3=5𝑒−6]. Middle: the restored images by Cai-TP. Bottom: the restored images by HADMM.

Example 2. We restore “Lena” contaminated by Gaussian noise of variance 𝜎²=25 plus impulse noise with different level. The blurred and restored images by MNID-ADMM are shown in Figure 4 and the results are shown in Table 2. These results show that our proposed method can restore corrupted image with various noise level efficiently and the restore quality is quite well.

Table 2 

The PSNR (dB) and computing time (seconds) for the restoration results by MNID-ADMM, Cai-TP, and HADMM for blurred Lena contaminated by Gaussian noise (𝜎=5) plus impulse noise. The blurring kernel is the out-of-focus kernel of radius 3.

Figure 4 

Top: Lena image blurred with out-of-focus kernel of radius 3 and then corrupted by Gaussian noise with 𝜎=5 and salt-and-pepper noise with 𝑠=30%, 50%, 70%, and 90%, respectively. The second line: results restored by MNID-ADMM (𝛼1=0.9,𝛽1=𝛽2=𝛽3=0.05). The third line: Lena image blurred with out-of-focus kernel of radius 3 and then corrupted by Gaussian noise with 𝜎=5 and random-valued noise with 𝑟=25%, 40%, 55%, and 70%, respectively. Bottom: results restored by MNID-ADMM (𝛼1=0.9, 𝛽1=𝛽2=𝛽3=0.05).

Example 3. We restore different images contaminated by Gaussian noise of variance 𝜎²=25 plus impulse noise with different level. The blurred and restored images by MNID-ADMM are shown in Figures 5 and 6, and the results are shown in Tables 3 and 4. From the experiments results, we conclude that the proposed method can restore different images with various noise level quite well with the same parameters.

Table 3 

The PSNR (dB) and computing time (seconds) for the restoration results by MNID-ADMM, Cai-TP, and HADMM for blurred images contaminated by Gaussian noise (𝜎=5) plus salt-and-pepper noise. The blurring kernel is the out-of-focus kernel of radius 3.

Table 4 

The PSNR (dB) and computing time (seconds) for the restoration results by MNID-ADMM, Cai-TP, and HADMM method for blurred images contaminated by random-valued noise plus Gaussian noise (𝜎=5). The blurring kernel is the out-of-focus kernel of radius 3.

Figure 5 

Top: images blurred with out-of-focus kernel of radius 3 and then corrupted by Gaussian noise with 𝜎=5 and salt-and-pepper noise with 𝑠=50%. The second line: results restored by MNID-ADMM (𝛼1=0.9, 𝛽1=𝛽2=𝛽3=0.05). The third line: Images blurred with out-of-focus kernel of radius 3, and then corrupted by Gaussian noise with 𝜎=5 and salt-and-pepper noise with 𝑠=70%. Bottom: results restored by MNID-ADMM (𝛼1=0.9, 𝛽1=𝛽2=𝛽3=0.05).

Figure 6 

Top: images blurred with out-of-focus kernel of radius 3 and then corrupted by Gaussian noise with 𝜎=5 and random-valued noise with 𝑟=55%. The second line: results restored by MNID-ADMM (𝛼1=0.9, 𝛽1=𝛽2=𝛽3=0.05). The third line: images blurred with out-of-focus kernel of radius 3, and then corrupted by Gaussian noise with 𝜎=5 and random-valued noise with 𝑟=70%. Bottom: results restored by MNID-ADMM (𝛼1=0.9, 𝛽1=𝛽2=𝛽3=0.05).

5. Conclusion

In this paper, a high-order total variation scheme is proposed to restore blurred images corrupted by Gaussian plus impulse noise. The alternating direction method of multipliers (ADMM) is employed to solve the proposed problem. Numerical examples demonstrate that the proposed model efficiently removes the mixed Gaussian plus impulse noise meanwhile avoids staircase effect. The proposed method can be extended to solve blind convolution problems. One of the difficulties in this paper is how to choose appropriate parameters and an adaptive parameter chosen method may be used in this method. These are future works.

Acknowledgments

This research is supported by 973 Program (2013CB329404), NSFC (61170311), and Sichuan Province Science and Technology Research Project (2012GZX0080).