
Citation: Myeongmin Kang, Miyoun Jung. Nonconvex fractional order total variation based image denoising model under mixed stripe and Gaussian noise[J]. AIMS Mathematics, 2024, 9(8): 21094-21124. doi: 10.3934/math.20241025
[1] | Miyoun Jung . A variational image denoising model under mixed Cauchy and Gaussian noise. AIMS Mathematics, 2022, 7(11): 19696-19726. doi: 10.3934/math.20221080 |
[2] | Donghong Zhao, Ruiying Huang, Li Feng . Proximity algorithms for the L1L2/TVα image denoising model. AIMS Mathematics, 2024, 9(6): 16643-16665. doi: 10.3934/math.2024807 |
[3] | Miyoun Jung . Group sparse representation and saturation-value total variation based color image denoising under multiplicative noise. AIMS Mathematics, 2024, 9(3): 6013-6040. doi: 10.3934/math.2024294 |
[4] | Abdelilah Hakim, Anouar Ben-Loghfyry . A total variable-order variation model for image denoising. AIMS Mathematics, 2019, 4(5): 1320-1335. doi: 10.3934/math.2019.5.1320 |
[5] | Lufeng Bai . A new approach for Cauchy noise removal. AIMS Mathematics, 2021, 6(9): 10296-10312. doi: 10.3934/math.2021596 |
[6] | Mingying Pan, Xiangchu Feng . Application of Fisher information to CMOS noise estimation. AIMS Mathematics, 2023, 8(6): 14522-14540. doi: 10.3934/math.2023742 |
[7] | Yating Zhu, Zixun Zeng, Zhong Chen, Deqiang Zhou, Jian Zou . Performance analysis of the convex non-convex total variation denoising model. AIMS Mathematics, 2024, 9(10): 29031-29052. doi: 10.3934/math.20241409 |
[8] | Hui Sun, Yangyang Lyu . Temporal Hölder continuity of the parabolic Anderson model driven by a class of time-independent Gaussian fields with rough initial conditions. AIMS Mathematics, 2024, 9(12): 34838-34862. doi: 10.3934/math.20241659 |
[9] | Yuzi Jin, Soobin Kwak, Seokjun Ham, Junseok Kim . A fast and efficient numerical algorithm for image segmentation and denoising. AIMS Mathematics, 2024, 9(2): 5015-5027. doi: 10.3934/math.2024243 |
[10] | Xiaodong Zhang, Junfeng Liu . Solving a class of high-order fractional stochastic heat equations with fractional noise. AIMS Mathematics, 2022, 7(6): 10625-10650. doi: 10.3934/math.2022593 |
Remote sensing images have been extensively used in various applications such as urban planning, military operations, and environment monitoring [1,2]. However, during the image acquisition process, remote sensing images are inevitably polluted by stripe noise, mainly due to the difference in the responses of the detectors and the calibration error. On the other hand, Gaussian noise is typically caused by the temperature of the sensor and the level of illumination in the environment that corrupts every pixel. The stripe noise is usually mixed with random Gaussian noise. This mixture of noise not only degrades the image quality but also hampers the subsequent processing such as classification [3], object segmentation [4], target detection [5], and image unmixing [6]. Therefore, removing this mixed noise is an essential preprocessing step for remote sensing images. In this work, we focus on restoring a remote sensing image in the presence of mixed stripes and Gaussian noise.
The removal of stripe noise in an image can be categorized as filtering-based methods, statistics-based methods, and optimization-based methods. Filtering-based methods eliminate stripe noise by truncating the stripe component in the transformed domain such as the Fourier transform [7], wavelet domain [8], and the combined domain [9]. These algorithms are simple but they assume that the stripe noise is periodic and identifiable in the power spectrum. Statistics-based methods such as moment matching [10] and histogram matching [11] assume that the statistical features of each sensor are the same. These methods have a low computational cost, but their performance is greatly affected by a predetermined reference moment or histogram.
Optimization-based models consider an ill-posed inverse problem for the stripe noise removal and utilize prior knowledge of the ideal image or the stripe noise as regularization [12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34]. For example, Bouali and Ladjal [13] proposed a unidirectional total variation (UTV) model for MODIS image destriping, by exploiting the direction feature of stripes. Several studies have developed the UTV model using various regularizations [14,15,16,17]. Despite the satisfactory performance of these UTV-based models, they directly restore the image without considering the characteristics of stripes, which generates a loss of image details along with the stripes. To overcome this shortcoming, various works make use of the structural property of stripes. In particular, in [18,24], the ℓ0-norm was used to constrain the global sparse distribution of stripes, but this assumption does not apply when the stripes are very dense. Chang et al. [19] proposed an image decomposition model by employing a low-rankness prior for stripes and TV regularization [35] for the image. They adopted the nuclear norm to express the global redundancy of stripes. Meanwhile, Yang et al. [27] exploited the Schatten 1/2-norm to characterize the low-rankness of stripes and unidirectional high-order TV for the image. In [20,22,25,34], the ℓ2,1-norm of stripes was suggested to promote the group sparsity of stripes. These destriping methods perform well when the stripes satisfy the low-rankness or group sparsity assumptions. However, these assumptions may be violated when the stripes are complex. To deal with complex stripes such as irregular or partial stripes, Wang et al. [26] suggested a reweighted ℓ2,1-norm regularization for stripes. However, this model only consider the features of stripes without prior image information, hence it cannot capture a clean image in the presence of Gaussian noise. In this work, we introduce a destriping model that enables restoring a clean image by simultaneously suppressing stripes and Gaussian noise, by utilizing the group sparsity characteristic and directional feature of stripes.
TV regularization has been widely used in various image restoration problems because of its edge-preserving advantage. However, TV tends to generate staircase artifacts in reconstructed images as it pursues piecewise-constant solutions. To mitigate the staircase effect, higher-order TV has been suggested; for example, second-order TV [36,37], total generalized variation [38], and hybrid TV [39,40]. Different from this type of high-order TV, fractional-order TV (FTV) uses derivatives with order greater than or equal to one, bringing a compact discrete form and thus yielding computational advantage. FTV regularization takes neighboring pixel values into account, so it preserves local geometric characteristics and thereby textures. Therefore, it has been adopted for various image processing problems, such as image denoising [41,42,43,44,45,46,47,48], texture enhancement [49], and super-resolution [50]. FTV has been empirically proven to suppress staircasing artifacts and improve the effectiveness of texture preservation. On the other hand, nonconvex regularization has attracted attention because nonconvex regularizers have advantages over convex regularizers in maintaining edges and details [51,52,53]. In various works [54,55,56,57], nonconvex higher-order TV has been developed, which contributes to edges conservation and the reduction of the staircase effect. Also, many efficient algorithms have been developed for solving nonconvex minimization problems. In particular, iteratively convex majorization-minimization methods for solving nonsmooth nonconvex minimization problems with convergence analysis have been proposed in [58], which are generalizations of the iteratively reweighted ℓ1 algorithm (IRL1) proposed for compressive sensing [59]. In the present work, we apply a nonconvex FTV regularziation to the image component to benefit from both FTV and nonconvex regularization. Besides, we employ IRL1 to solve the proposed model, along with a convergence analysis.
In this article, we introduce a novel image denoising model in the presence of a mixture of stripe and Gaussian noises. In this work, we consider a relatively high level of Gaussian noise, unlike previous works that consider a low level of Gaussian noise. To effectively remove the mixed noise and recover a clean image, the proposed model exploits prior knowledge of both stripe noise and image components. In particular, the group sparsity feature and directional property of stripes are utilized to extract stripe noise. Besides, a nonconvex FTV is used for the image component to recover its smooth regions with less staircase artifacts while preserving edges. To solve the proposed nonconvex model, we employ the IRL1 algorithm and alternating direction method of multipliers [60,61,62] and provide a convergence analysis.
The remainder of this article is organized as follows. In Section 2, we recall several optimization-based destriping models and review the FTV. Section 3 introduces the proposed model for the removal of stripes and Gaussian noise. An optimization algorithm for solving the proposed model is also provided, and its convergence is proven. Section 4 presents the experimental results of the proposed model, comparing it with several existing models. Finally, in Section 5, we summarize our work and provide some remarks.
This section reviews several optimization-based models for stripe noise removal. In remote sensing images, stripe noise typically includes additive and multiplicative noise components [12]. The multiplicative noise can be described as additive noise by a logarithmic operation [63], thus stripe noise can be regarded as additive noise. Therefore, the degradation model for the removal of stripe noise is usually given by
f=u+s+n, |
where f, u, s, and n:Ω→R, where Ω={(x1,x2):x1=1,2,⋯,M,x2=1,2,⋯,N} (M and N denote the number of columns and rows of the 2D gray-scale image, respectively), represent the observed image, the desired clear image, stipe noise, and additive Gaussian white noise, respectively. Stripes are generally assumed to be vertical (x2-direction). If the stripes are horizontal, one can rotate them to make the stripes vertical.
First, Chang et al. [19] introduced an image decomposition model that simultaneously models the characteristics of stripe and image components. Specifically, they utilized a low-rank constraint for stripes and TV regularization for the image, which led to the following model:
minu,s12‖f−u−s‖22+λ1‖∇u‖1+λ2‖s‖∗, |
where λi (i=1,2) are regularization parameters. Here, ‖∇u‖1 is the anisotropic TV such as ‖∇u‖1=‖∇x1u‖1+‖∇x2u‖1, where ∇u=(∇x1u,∇x2u)T with ∇x1 and ∇x2 denoting the horizontal and vertical derivative operators, respectively. ‖A‖∗ represents the nuclear norm of the matrix A, defined as the sum of its singular values, i.e., ‖A‖∗=∑iσi(A). This model performs well for handling vertical stripes, but is not explicitly applicable to oblique stripes.
Wang et al. [25] proposed a novel unified destriping model to effectively exploit the low-rankness and group sparsity of the oblique stripe noise:
minu,s12‖f−u−s‖22+λ1‖∇x1u‖1+λ2‖τ∘s‖∗+λ3‖τ∘s‖2,1, |
where λi>0 (i=1,2,3) are parameters, and τ is the shear operator to transform the oblique stripe noise vertically, which does not include rotation and filling up. ‖s‖2,1=∑i‖[s]i‖2 represents the group sparsity, where [s]i is the i-th column of s. This model attains excellent performance on thin and regular stripes, but it cannot effectively remove agglomerated, banded, and irregular stripes while generating the staircase effect in restored images.
To overcome these drawbacks, the authors in [27] utilized a unidirectional higher-order TV regularization for the image, and the Schatten 1/2-norm to characterize the low-rankness of stripes:
minu,s12‖f−u−s‖22+λ1‖∇x1u‖1+λ2‖∇2x1x1u‖1+λ3‖s‖1/2s1/2, |
where λi>0 (i=1,2,3) are parameters, ∇2x1x1 is the second-order gradient operator across the horizontal direction, and ‖⋅‖1/2s1/2 is the Schatten 1/2-norm defined as ‖s‖1/2s1/2=∑iσ1/2i with all singular values σi of s. Despite the phenomenal destriping performance of the aforementioned models, the low-rank prior for stripe noise may be violated in remote sensing images, such as the stripes with small fragment cases.
On the other hand, there are various works [20,21,23,24] that utilize the directional property of stripes, i.e., unidirectional gradient sparsity regularization of s. Among them, the authors in [20] proposed an image decomposition model that can handle both stripes and Gaussian noise:
minu,s12‖f−u−s‖22+λ1‖∇x1u‖1+λ2‖∇x2u‖1+λ3‖∇x2s‖1+λ4‖s‖2,1, |
where λi>0 (i=1,2,3,4) are parameters. The term ‖∇x2s‖1 enforces that the stripe component has good smoothness in vertical direction.
This subsection recalls fractional-order derivatives and FTV. Fractional-order derivatives are seen as a generalization of the integer-order derivatives. Three well-known definitions of fractional-order derivatives are the Riemann-Liouville, Grünwald-Letnikov, and Caputo definitions [64,65,66]. In particular, the Grünwald–Letnikov (GL) fractional-order derivative is based on the finite difference method and is easy to implement. For one-dimensional signals f(x), x∈[a,x], the GL fractional α-order derivative is defined as
Dαf(x)=dαf(x)dxα=limh→01hα[x−ah]∑j=0(−1)j(αj)f(x−jh), |
where α>0, [b] is the integer such that b−1<[b]≤b, and (αj)=α(α−1)⋯(α−j+1)j! is the combination parameter.
For a function u:Ω→R, where Ω⊂R2 is an open and bounded set with compact support, let ∇αxiu=∂αu∂xαi (i=1,2) be the fractional α-order derivative Dαu of u along the xi-direction. Then, the anisotropic fractional α-order TV is defined as
∫Ω∣∇αu∣dx=∫Ω∣∇αx1u∣+∣∇αx2u∣dx, | (2.1) |
while its isotropic version is given by ∫Ω√(∇αx1u)2+(∇αx2u)2dx.
If Ω={(x1,x2):x1=1,2,⋯,M,x2=1,2,⋯,N} is a discretized domain, then an image u defined on Ω can be represented as a matrix in RN×M, and ui,j denotes the (i,j)-th element of u (i=1,...,M, j=1,...,N). Then, its discrete fractional-order derivatives ∇αx1u and ∇αx2u are given by
(∇αx1u)i,j=K−1∑k=0(−1)kCαkui−k,j,(∇αx2u)i,j=K−1∑k=0(−1)kCαkui,j−k, |
where K is the number of neighboring pixels used in computation of the fractional-order derivatives at each pixel, and Cαk denotes the generalized binomial coefficients as Cαk=Γ(α+1)Γ(k+1)Γ(α−k+1), with Γ(⋅) denoting the Gamma function. For fixed α, the coefficients Cαk rapidly tend to zero as k increases. Then, the discretized version of FTV in (2.1) is given by ‖∇αu‖1=∑i,j∣(∇αx1u)i,j∣+∣(∇αx2u)i,j∣. According to [43], the high-pass capability becomes stronger with larger α, so more texture regions are preserved when α increases. The experimental results in the literature [41,42,43,44,45,46,47,48] show that FTV performs well in terms of removing the staircase effect while preserving textures.
In this section, we introduce a denoising model to remove both stripe noise and Gaussian noise. We also present an optimization algorithm for solving the proposed model.
We assume that the observed noisy image f:Ω→R is degraded by both stripe noise and Gaussian noise as follows:
f(x1,x2)=u(x1,x2)+s(x1,x2)+n(x1,x2), | (3.1) |
where Ω={(x1,x2):x1=1,2,⋯,M,x2=1,2,⋯,N}. Here, u is the clean image, s represents the periodic or non-periodic stripes that are vertical (x2-direction), and n represents the Gaussian noise following the normal distribution, N(0,σ2), with a standard deviation σ.
In this work, unlike in previous works, the Gaussian noise level is considered to be relatively high, hence we intend to restore the image u by eliminating both stripe noise and Gaussian noise simultaneously. To effectively retrieve u from the data f in (3.1), we propose the following image decomposition model:
minu,s12‖f−u−s‖22+λ1⟨ϕ(∣∇αx1u∣),1⟩+λ2⟨ϕ(∣∇αx2u∣),1⟩+λ3‖∇x2s‖1+λ4∑x1ψ(‖s(x1,⋅)‖2), | (3.2) |
where ⟨⋅,⋅⟩ denotes the inner product, λi>0 (i=1,2,3,4) are regularization parameters, 1<α<2, and ‖s(x1,⋅)‖2=√∑x2s(x1,x2)2. Moreover, ϕ and ψ are given by the following nonconvex functions
ϕ(v)=1ρlog(1+ρv),ψ(w)=log(β+w), |
where ρ>0 controls the nonconvexity of FTV, and β>0 is a small parameter.
The first term in (3.2) is a data-fidelity term that estimates the discrepancy between f and u+s. The second and third terms control the smoothness of u, which is a nonconvex version of the anisotropic FTV. This FTV alleviates the staircase effect that is commonly seen in restored images from TV-based models. The nonconvex function ϕ does not penalize the formulation with strong gradients of u, thus protecting large details and textures in the image. On the other hand, at near-zero points (v→0+), it is preferable for ϕ(v) to have the same behavior as the linear function, v, so that u can be better smoothed in homogeneous regions of the image. That is, the nonconvex function ϕ further enforces the preservation of edges or discontinuties, so the use of the nonconvex FTV (NFTV) leads to higher PSNR and SSIM values than using FTV, as shown in Figure 1. In fact, there are more choices for the nonconvex function ϕ such as vq (0<q<1) and v1+ρv. However, it is hard to solve the minimization problem involving the nonconvex q-norm regularizer because finding a limiting-supergradient of ‖⋅‖q at zero is difficult. Besides, the fractional nonconvex function is more proper for the reconstruction of piecewise-constant images since limv→∞ϕ(v)=c (c is constant), while the logarithmic function is suitable for the reconstruction of images that are no longer piecewise-constant since limv→∞ϕ(v)=∞, as explained in [67]. There are many real synthetic aperture radar (SAR) images that are not piecewise-constant; thus, we utilize the logarithmic function as our nonconvex function ϕ. The regularization parameters λ1 and λ2 may be different. For example, when the Gaussian noise level is low, λ1 may be chosen to be larger than λ2. However, as the Gaussian noise level increases, smoothing of u along the x2-direction is also necessary, so λ2 must be close to λ1. In practice, we consider two Gaussian noise levels, σ=10 or 20, which are relatively high, so we set the values of λ1 and λ2 be the same.
The last two terms in (3.2) exploit the directional characteristics and group sparsity of stripe noise. In particular, the fourth term is a unidirectional TV regularization of the stripe component that imposes the sparsity of its vertical derivatives. Furthermore, the stripe component is composed of stripe lines and stripe-free lines, and each line can be viewed as a group. Thus, ∑x1‖s(x1,⋅)‖2 enforces the group sparsity (GS) of stripes. However, ℓ2,1-norm is not able to effectively promote the group sparsity of stripes in many cases [26,59]. As a result, we characterize the intrinsic structure of stripes using a nonconvex function ψ for ‖s(x1,⋅)‖2. Likewise, wp (0<p<1) or wβ+w could be other options for the nonconvex function ψ. Figure 1 shows a comparison of the proposed model (3.2) with a model (3.2) that includes the GS term, ∑x1‖s(x1,⋅)‖2, instead of the nonconvex GS (NGS) term. The model (3.2) with the GS term fails to properly extract the stripe noise, resulting in some traces of stripes in the restored images. This can also be seen more clearly in the difference images in the second and forth rows of Figure 1. Meanwhile, the proposed model (3.2), which includes the NGS term, more suitably extracts the stripe noise from the images. This also leads to the restored images with better conserved details and edges.
In this section, we present an optimization algorithm for solving the proposed model in (3.2). Given the matrix f∈RN×M, model (3.2) can be rewritten as
minu,s12‖f−u−s‖22+λ1⟨ϕ(∣∇αx1u∣),1⟩+λ2⟨ϕ(∣∇αx2u∣),1⟩+λ3‖∇x2s‖1+λ4∑ilog(β+‖[s]i‖2), | (3.3) |
where [s]i is the i-th column of s, with i=1,2,...M.
To solve the nonconvex problem (3.3), we first employ the IRL1 proposed in [58] for solving a nonconvex minimization problem. Let us consider the following nonconvex minimization problem: minzE1(z)+E2(G(z)), where E1 is a proper, lower semicontinuous and convex function, E2 is a concave and coordinatewise nondecreasing function, and G is a coordinatewise convex function. Applying IRL1 to this problem leads to the following iterative algorithm:
{wℓ+1∈ˉ∂E2(G(zℓ)),zℓ+1:=argminzE1(z)+⟨wℓ+1,G(z)⟩, | (3.4) |
where ˉ∂E2:=−∂(−E2) denotes the superdifferential of the function E2. For the global convergence of IRL1, it is required that E1(z)+⟨wℓ+1,G(z)⟩ is strongly convex with a constant independent of ℓ. Thus, the authors in [58] suggest a modified version of IRL1 by adding a proximal term δ2‖z−zℓ‖22 to the convex surrogate problem in (3.4) with arbitrarily small δ>0.
To apply IRL1 to model (3.3), we can set up as follows:
E1(u,s)=12‖f−u−s‖22+λ3‖∇x2s‖1,E2(v1,v2,t)=λ1⟨ϕ(v1),1⟩+λ2⟨ϕ(v2),1⟩+λ4∑ilog(β+ti),G(u,s)=(∣∇αx1u∣,∣∇αx2u∣,‖[s]1‖2,‖[s]2‖2,...,‖[s]M‖2), | (3.5) |
where t=(t1,t2,...,tM).
Then, we adopt the modified IRL1 in [58], which is obtained by adding two proximal terms, δ2‖u−uℓ‖22 and δ2‖s−sℓ‖22:
{wℓ+11=11+ρ∣∇αx1uℓ∣,wℓ+12=11+ρ∣∇αx2uℓ∣,(wℓ+13)i=1β+‖[sℓ]i‖2,i=1,⋯,M(uℓ+1,sℓ+1):=argminu,s12‖f−u−s‖22+λ3‖∇x2s‖1+λ1⟨wℓ+11,∣∇αx1u∣⟩+λ2⟨wℓ+12,∣∇αx2u∣⟩+λ4∑i(wℓ+13)i‖[s]i‖2+δ2‖u−uℓ‖22+δ2‖s−sℓ‖22, | (3.6) |
where (λ1w1,λ2w2,λ4w3)T∈ˉ∂E2(G(u,s)) with ˉ∂E2=∇E2, and δ>0 is a small parameter.
Based on the convergence of IRL1 in [58], we prove the global convergence of IRL1 (3.6) as follows:
Theorem 1. Let {(uℓ,sℓ)} be the sequence generated by the IRL1 (3.6). Then, {(uℓ,sℓ)} converges to (u∗,s∗) as ℓ→∞, where (u∗,s∗) is a critical point of (3.3). Furthermore, the sequence {(uℓ,sℓ)} has finite length: ∑∞ℓ=0‖uℓ−uℓ+1‖2+‖sℓ−sℓ+1‖2<∞.
Proof. We need to check the assumptions in Theorem 2 in [58]. First, the objective function in (3.3) (or E(u,s)=E1(u,s)+E2(G(u,s)) from (3.5)) is clearly coercive and bounded below. In addition, due to the proximal terms, δ2‖u−uℓ‖22 and δ2‖s−sℓ‖22, the objective function of the convex subproblem in (3.6) is strongly convex with a constant independent of ℓ. Next, we show that the following three assumptions are satisfied:
(a) The objective function E(u,s) has the Kurdyka–Łojasiewicz (KL) property at a cluster point.
(b) E2(v1,v2,t) has locally Lipschitz continuous gradients on a compact set containing all the points G(uℓ,sℓ).
(c) The convex function z↦⟨wℓ+1,z⟩ for all ℓ, where z=(v1,v2,t)T and w=(w1,w2,w3)T, has a globally Lipschitz continuous gradient with a common Lipschitz constant.
a) A function is called a KL function if the function is lower semicontinuous and KL inequality holds for every point in the domain. According to [68], polynomials, indicator function, ‖⋅‖1, and ‖⋅‖2 are KL functions. Moreover, the log and exponential functions are also KL functions. Indeed, they are included in the log–exp structure [69], and the functions that are definable in such an o-minimal structure have the KL property. Hence, the objective function in (3.3) is a KL function.
b) The gradient and Hessian of E2 are given by
∇E2(v1,v2,t)=(λ11+ρv1,λ21+ρv2,λ4β+t1,...,λ4β+tM)T,∇2E2(v1,v2,t)=[−λ1ρ(1+ρv1)200⋯00−λ2ρ(1+ρv2)20⋯000−λ4(β+t1)2⋯0⋮⋮⋮⋱⋮000⋯−λ4(β+tM)2]. |
Hence, ‖∇2E2‖∞≤max(λ1ρ,λ2ρ,λ4/β2) on the image of G. Thus, E2 has a Lipschitz continuous gradient.
c) Trivially, z↦⟨wℓ+1,z⟩ has a globally Lipschitz continuous gradient with a common Lipschitz constant 0.
Therefore, all the assumptions of Theorem 2 in [58] are satisfied, so the theorem is proved.
Now we solve the (u,s)-subproblem in IRL1 (3.6). This subproblem is convex, but it involves non-differentiable terms. To resolve this problem, numerous efficient convex optimization algorithms have been suggested, such as [60,61,62,70,71]. In particular, we adopt the alternating direction method of multipliers (ADMM) in [60,61,62]. The ADMM is a widely-known algorithm for solving linearly constrained convex optimization problems, with its convergence proven in [61,62].
First, we introduce auxiliary variables pi (i=1,2,3,4), based on the variable splitting technique. Hence, the (u,s)-subproblem in (3.6) can be converted into the following constrained minimization problem:
minu,s,p1,p2,p3,p412‖f−u−s‖22+λ1⟨wℓ+11,∣p1∣⟩+λ2⟨wℓ+12,∣p2∣⟩+λ3‖p3‖1+λ4∑i(wℓ+13)i‖[p4]i‖2+δ2‖u−uℓ‖22+δ2‖s−sℓ‖22,subject to:p1=∇αx1u,p2=∇αx2u,p3=∇x2s, p4=s. | (3.7) |
The augmented Lagrangian function of problem (3.7) is given by
Lμ(u,s,→p,→h)=12‖f−u−s‖22+λ1⟨wℓ+11,∣p1∣⟩+λ2⟨wℓ+12,∣p2∣⟩+λ3‖p3‖1+λ4∑i(wℓ+13)i‖[p4]i‖2+δ2‖u−uℓ‖22+δ2‖s−sℓ‖22−⟨h1,p1−∇αx1u⟩+μ2‖p1−∇αx1u‖22−⟨h2,p2−∇αx2u⟩+μ2‖p2−∇αx2u‖22−⟨h3,p3−∇x2s⟩+μ2‖p3−∇x2s‖22−⟨h4,p4−s⟩+μ2‖p4−s‖22, |
where →p=(p1,p2,p3,p4)T, →h=(h1,h2,h3,h4)T, where hi∈RN×M×2 (i=1,2,3) and h4∈RN×M are the Lagrangian multipliers, and μ>0 is a penalty parameter.
Then, the ADMM applied to (3.7) brings the following iterative algorithm:
{(uk+1,sk+1):=argminu,sLμ(u,s,→pk,→hk)→pk+1:=argmin→pLμ(uk+1,sk+1,→p,→hk),hk+11=hk1−γμ(pk+11−∇αx1uk+1),hk+12=hk2−γμ(pk+12−∇αx2uk+1),hk+13=hk3−γμ(pk+13−∇x2sk+1),hk+14=hk4−γμ(pk+14−sk+1), | (3.8) |
where γ∈(0,√5+12). We can attain the following convergence results according to the convergence in [61]:
Theorem 2. If the sequence {(uk,sk,→pk,→hk)} is generated by ADMM in (3.8) and γ∈(0,√5+12), then {(uk,sk,→pk)} strongly converges to a limit point (u∗,s∗,→p∗), {→hk+1−→hk} converges to 0, and {→hk} is bounded. Moreover, if →h∗ is a weak cluster point of →hk, then (u∗,s∗,→p∗,→h∗) is a saddle point of the augmented Lagrangian Lμ.
Now we solve the (u,s)-subproblem in ADMM (3.8). This subproblem can be reformulated as the following least squares problem:
(uk+1,sk+1):=argminu,s12‖f−u−s‖22+δ2‖u−uℓ‖22+δ2‖s−sℓ‖22+μ2‖∇αx1u−pk1+hk1/μ‖22+μ2‖∇αx2u−pk2+hk2/μ‖22+μ2‖∇x2s−pk3+hk3/μ‖22+μ2‖s−pk4+hk4/μ‖22. |
The first-order optimality condition leads to the following normal equation:
[B1IIB2][us]=[RHSuRHSs], | (3.9) |
where B1, B2, RHSu, and RHSs are given by
B1=(1+δ)I+μ(∇αx1)T∇αx1+μ(∇αx2)T∇αx2,B2=(1+δ+μ)I+μ(∇x2)T∇x2,RHSu=f+δuℓ+μ(∇αx1)T(pk1−hk1/μ)+μ(∇αx2)T(pk2−hk2/μ),RHSs=f+δsℓ+μ(∇x2)T(pk3−hk3/μ)+μ(pk4−hk4/μ). |
Here, (∇α)T=(−1)αdivα, where divαq∈RN×M for q=(q1,q2)∈RN×M×2 is the discrete fractional-order divergence defined as
(divαq)i,j=(∇αx1q1)i,j+(∇αx2q2)i,j. |
The elements B1, B2, and I in Eq (3.9) can be diagonalized by the 2-dimensional fast Fourier transform (FFT2) under the periodic boundary condition. Thus, the block matrix in the left-hand side of Eq (3.9) can be diagonalized by using FFT2. Therefore, the solution (uℓ+1,sℓ+1) can be obtained exactly using the inversion formula of the block matrix.
Next, we solve the →p-subproblem in ADMM (3.8). The variables pi are independent of each other, so we can solve the subproblem for each pi:
pk+11:=argminp1λ1⟨wℓ+11,∣p1∣⟩+μ2‖p1−∇αx1uk+1−hk1/μ‖22,pk+12:=argminp2λ2⟨wℓ+12,∣p2∣⟩+μ2‖p2−∇αx2uk+1−hk2/μ‖22,pk+13:=argminp3λ3‖p3‖1+μ2‖p3−∇x2sk+1−hk3/μ‖22,pk+14:=argminp4λ4∑i(wℓ+13)i‖[p4]i‖2+μ2‖p4−sk+1−hk4/μ‖22. | (3.10) |
The p1-subproblem in (3.10) has the closed form solution as
pk+11=shrink(∇αx1uk+1+hk1/μ,λ1wℓ+11/μ), | (3.11) |
where shrink is the soft-thresholding operator defined as
shrink(a,b)t=at∣at∣⋅max(∣at∣−bt,0),t∈Ω. |
Similarly, pk+12, pk+13, and pk+14 are explicitly achieved as
pk+12=shrink(∇αx2uk+1+hk2/μ,λ2wℓ+12/μ),pk+13=shrink(∇x2sk+1+hk3/μ,λ3/μ),[pk+14]i=[˜pk4]i‖[˜pk4]i‖2⋅max(‖[˜pk4]i‖2−λ4(wℓ+13)i/μ,0), | (3.12) |
with ˜pk4=sk+1+hk4/μ and i=1,...,M.
Consequently, the proposed algorithm is summarized in Algorithm 1.
Algorithm 1 IRL1 for solving model (3.2). |
1: Input: choose the parameters λi (i=1,2,3,4), α, β, ρ, δ, μ>0, γ∈(0,√5+12) and the maximum numbers of iterations Nout, Nin. |
2: Initialization: u0=f, s0=0, p0i=0, h0i=0 (i=1,2,3,4). |
3: for k=0,1,2,⋯Nout do |
4: Compute wℓ+1i (i=1,2,3) using (3.6), |
5: for k=0,1,2,⋯,Nin do |
6: Compute (uk+1,sk+1) by solving Eq (3.9) using FFT2, |
7: Compute pk+1i (i=1,2,3,4) using (3.11) and (3.12), |
8: Update hk+1i (i=1,2,3,4) using (3.8), |
9: end for |
10: end for |
11: Output: restored image u. |
This section presents numerical results for the removal of mixed stripe and Gaussian noise. We compare the performance of the proposed model with existing image decomposition models such as LRSID [19], LRGS [25], TVGS [20], Schatten [27], and ELRTV [30]. In the LRGS, Schatten, and ELRTV models, they use a unidirectional TV for the image component and thus fail to adequately remove the Gaussian noise in the presence of a high level of Gaussian noise. Thus, after removing the stripe noise by adopting these models, we utilize the TV denoising model [35] as a postprocessing step to remove the Gaussian noise. We call these models LRGS+TV, Schatten+TV, and ELRTV+TV, respectively. All numerical results are available in the material at the following link: https://han.gl/ouTofQ.
The original remote sensing images are given in Figure 2, and the range of intensity values in the original images is assumed to be [0,1]. We consider two types of stripe noise such as periodic or non-periodic stripe noise and assume that the stripes are vertical. Specifically, we randomly select columns of the image to add stripes. In the case of periodic stripes, initial stripes are randomly selected from the first 32 (period = 32) columns, and these stripes are periodically added to the original images. On the other hand, non-periodic stripes are randomly selected from the entire columns. The amount of stripe noise is determined by the percentage of degraded region, r, and the intensity of the added stripes, m. In our experiments, we select r∈{30,50,70} and m∈{50,100}. Moreover, the Gaussian noise level, σ, is set to 10 or 20. The numerical experiments were implemented using MATLAB R2020b on a 64-bit Windows 10 operating system using an Intel Xeon Silver CPU at 2.40 GHz and 64 GB memory.
To estimate the quality of the restored images, we compute the peak-signal-to-noise-ratio (PSNR) value, which is defined as
PSNR(u,u∗)=10log10(MN‖u−u∗‖22), |
where u and u∗ represent the recovered image and the original image, respectively, and MN is the size of the image. We also calculate the structure similarity (SSIM) index [72], which is a perception-based measure that carries visual information about the structure of the objects.
The stopping criterion of the proposed model is given by
‖uℓ−uℓ−1‖2‖uℓ‖2<tol or ℓ>MaxIter, |
where tol is a given tolerance number, and MaxIter is a given maximum iteration number. For our outer loop, we set tol=10−4 and Nout=400, and for our inner loop, Nin=1. We use the stopping conditions given in their own works for existing models.
The parameters are tuned to achieve the best visual quality of restored image. The parameter settings of the proposed model are as follows. First, we set the same values for the regularization parameters λ1 and λ2, which depend on the level of Gaussian noise. When σ=10, λ1 and λ2 are chosen from {0.02,0.03}, and when σ=20, they are chosen from {0.04,0.05,0.06}. Meanwhile, the parameter λ3 and λ4 are the regularization parameters for the stripe noise component. λ3 is fixed at 0.6, while λ4 is chosen more carefully than λ3. Specifically, for periodic stripes, λ4∈{0.05,0.08,0.1,0.2,0.4,0.6,0.8}, while for non-periodic strips, λ4∈{0.01,0.03,0.05,0.08,0.1,0.2,0.4} when r=30 or 50, and λ4∈{0.005,0.01,0.03,0.05,0.08} when r=70. The values of λ1 and λ4 are presented in each figure. The parameter β in the NGS term is set to 10−15, and the parameter ρ in the NFTV terms is set to 1. The derivative order α in the FTV terms is set to 1.3 or 1.5, and K is set to 20. The parameter δ in IRL1 is fixed at 0.0001. The parameters μ and γ in the ADMM algorithm are set to 0.1 and 1.618, respectively.
In this section, we present the denoising results in the presence of periodic stripe noise and Gaussian noise.
First, Figure 3 presents the data images of a natural image in the presence of periodic stripe noise with r=50 or 70 and m=100 and Gaussian noise with σ=20, while Figures 4 and 5 present the denoising results. The difference images between the restored and original images are also presented to effectively show the denoising results. First, it can be seen that LRSID, LRGS+TV, Schatten+TV, and ELRTV+TV models fail to correctly decompose the stripe and image components. This leads to the leftover of some stripes or loss of details in the restored images, which are clearly visible in the difference images. Meanwhile, our model and TVGS better extract the stripe noise than the aforementioned models that utilize the low-rankness of stripes. This is due to the use of a directional term and a group sparsity term for the stripes. Comparing our model with TVGS, their restored images look very similar, but we can see from the difference images that our model removes the stripe noise better than TVGS. This shows the effectiveness of our nonconvex group sparsity term for stripes extraction. Moreover, our FTV regularization helps mitigate the staircase artifacts found in the restored images of TVGS. Besides, our nonconvex FTV regularization allows the conservation of finer features and details, so the difference images of our model include much fewer image structures than other models. All these observations result in higher PSNR and SSIM values for our model than TVGS. As a result, these examples show better denoising performance of the proposed model than other models, by effectively removing both stripes and Gaussian noise in a natural image. Since we focus on denoising of SAR images in this work, we put more denoising results for natural images in the supplementary file at the following link: https://url.kr/6wj8n7.
Figures 6 and 7 present the denoising results tested on real SAR images in the presence of periodic stripe noise with r=30, 50, or 70 and m=50 and Gaussian noise with σ=20. Similar to the previous results, the LRSID, LRGS+TV, Schatten+TV, and ELRTV+TV models struggle to properly extract stripes from the images compared with our model and TVGS. Also, while TVGS appears to provide similar restored images to our model, our model separates the stripe and image components better than TVGS, which can be seen more clearly in the difference images. Thus, these also show the effective denoising performance of the proposed model for SAR images.
In Figure 8, we present the denoising results of our model with different values of m (m=50 or 100), while r=50 and σ=20 are fixed, and we compare our model with TVGS. We can see that our model separates stripes with image structures better than TVGS and that our difference images have much fewer streaks and image edges. On the other hand, despite the use of different intensity values for the stripes, both models supply similar visual quality in the restored images, leading to similar PSNR and SSIM values. Indeed, throughout the experiments, the denoising results of our model are similar when m=50 and m=100, in terms of visual quality of restored images and PSNR and SSIM values.
Table 1 presents the mean PSNR and SSIM values of all methods tested on all images in Figure 2, in the presence of periodic stripe noise and Gaussian noise. The PSNR and SSIM values for all image cases are given in the material at the following link: https://han.gl/ouTofQ. As the noise levels σ or r increase,the PSNR and SSIM values decrease. In all cases,the proposed model provides the highest average PSNR and SSIM values. This also verifies the superior denoising performance of our model over other models when both periodic stripes and Gaussian noise exist.
r | 30 | 50 | 70 | ||||
m | 50 | 100 | 50 | 100 | 50 | 100 | |
σ=10 | LRSID | 28.20/0.8760 | 27.27/0.8640 | 27.86/0.8733 | 27.35/0.8665 | 28.02/0.8749 | 27.41/0.8664 |
TVGS | 29.06/0.8849 | 29.02/0.8848 | 28.78/0.8809 | 28.74/0.8807 | 28.59/0.8784 | 28.51/0.8775 | |
LRGS | 26.94/0.8075 | 26.93/0.8073 | 26.86/0.8070 | 26.84/0.8071 | 26.72/0.8062 | 26.70/0.8060 | |
LRGS+TV | 27.09/0.8599 | 27.09/0.8600 | 27.01/0.8588 | 26.98/0.8586 | 26.85/0.8575 | 26.83/0.8574 | |
Shatten | 26.69/0.7945 | 26.31/0.7919 | 26.43/0.7920 | 26.59/0.7943 | 26.56/0.7941 | 26.44/0.7923 | |
Shatten+TV | 27.51/0.8688 | 27.04/0.8656 | 27.09/0.8594 | 27.37/0.8678 | 27.34/0.7941 | 27.11/0.7923 | |
ELRTV | 25.73/0.7918 | 25.51/0.7896 | 25.77/0.7923 | 25.74/0.7921 | 25.78/0.7923 | 25.75/0.7921 | |
ELRTV+TV | 25.94/0.8456 | 25.71/0.8433 | 25.99/0.8458 | 25.96/0.8455 | 25.99/0.8459 | 25.97/0.8458 | |
Our | 29.26/0.8878 | 29.26/ 0.8879 | 29.06/0.8847 | 29.07/0.8849 | 28.90/ 0.8825 | 28.87/0.8822 | |
σ=20 | LRSID | 25.01/0.7595 | 24.43/0.7480 | 24.92/0.7595 | 24.50/0.7525 | 25.05/0.7614 | 24.65/0.7552 |
TVGS | 25.60/0.7783 | 25.57/0.7784 | 25.42/0.7742 | 25.38/0.7732 | 25.28/0.7702 | 25.21/0.7695 | |
LRGS | 22.33/0.6149 | 22.31/0.6147 | 22.29/0.6148 | 22.29/0.6146 | 22.24/0.6145 | 22.23/0.6139 | |
LRGS+TV | 24.83/0.7625 | 24.81/0.7630 | 24.75/0.7619 | 24.76/0.7619 | 24.67/0.7607 | 24.66/0.7608 | |
Shatten | 21.86/0.6022 | 21.73/0.6001 | 21.84/0.6024 | 21.80/0.6010 | 21.83/0.6016 | 21.81/0.6015 | |
Shatten+TV | 24.72/0.7564 | 24.49/0.7547 | 24.67/0.7566 | 24.63/0.7566 | 24.71/0.7570 | 24.64/0.7566 | |
ELRTV | 21.92/0.6064 | 21.81/0.6047 | 21.93/0.6066 | 21.91/0.6060 | 21.93/0.6066 | 21.90/0.6062 | |
ELRTV+TV | 24.29/0.7572 | 24.11/0.7551 | 24.30/0.7573 | 24.27/0.7567 | 24.30/0.7569 | 24.27/0.7570 | |
Our | 25.79/0.7861 | 25.78/0.7864 | 25.63/0.7820 | 25.62/0.7817 | 25.41/0.7766 | 25.45/0.7775 |
In Figure 9, we present the denoising results of our model at different periods, such as P=16, 32, and 64. We can observe that our model provides similar denoising performance despite the change of P. In Table 2, we present the PSNR and SSIM values of all models tested on three different images, when P=16, 32, and 64. In the case of LRGS+TV, P=32 provides higher PSNR values than other cases. LRSID and TVGS supply the lowest PSNR values when P=64, whereas Schtten+TV and ELRTV+TV provide the lowest PSNR values when P=16. However, our model provides similar PSNR and SSIM values for different values of P, and our PSNR values are higher than other models for all cases. This indicates that our model is not sensitive to the value of P in contrast to other models.
Image | P | LRSID | TVGS | LRGS+TV | Shatten+TV | ELRTV+TV | Our |
(a) | 16 | 24.80/0.6388 | 26.72/0.6921 | 25.69/0.6921 | 24.81/0.6635 | 24.81/0.6774 | 27.15/0.7072 |
32 | 25.38/0.6660 | 26.43/0.6919 | 25.99/0.6966 | 25.69/0.6843 | 25.69/0.6984 | 27.07/0.7088 | |
64 | 23.54/0.6157 | 26.32/0.6628 | 25.81/0.6876 | 25.60/0.6801 | 25.57/0.6905 | 27.04/0.7066 | |
(b) | 16 | 25.35/0.6999 | 26.45/0.6963 | 26.34/0.7034 | 25.08/0.6797 | 25.09/0.7084 | 26.77/0.7032 |
32 | 26.23/0.7042 | 26.41/0.6944 | 26.64/0.7038 | 25.83/0.6826 | 26.04/0.7109 | 26.70/0.7028 | |
64 | 23.10/0.6851 | 26.00/0.6926 | 26.32/0.7031 | 25.95/0.6834 | 25.76/0.7096 | 26.70/0.7029 | |
(g) | 16 | 23.91/0.7935 | 25.05/0.8025 | 24.67/0.8044 | 24.04/0.7968 | 24.09/0.8066 | 25.24/0.8095 |
32 | 24.00/0.7805 | 25.01/0.8014 | 24.93/0.7780 | 24.69/0.7972 | 24.81/0.8069 | 25.20/0.8076 | |
64 | 21.40/0.7714 | 24.92/0.8016 | 24.85/0.8045 | 24.70/0.7982 | 24.77/0.8082 | 25.21/0.8092 |
Lastly, Figure 10 depicts the impact of parameters λ1, α, λ3, and λ4. As mentioned earlier, we set the same values for λ1 and λ2. First, λ1 and α control the smoothness of the recovered images. Specifically, as the value of λ1 increases, the restored image becomes smoother. Although λ1=0.03 provides the highest PSNR value, the restored image with λ1=0.03 retains some Gaussian noise. Thus, we choose the restored image with λ1=0.04 as the best image. In the whole test, we selected the best restored images, considering both their visual quality and PSNR and SSIM values. Second, as the value of α increases, the staircase effect, which occurs when α=1, becomes alleviated, but the restored images becomes smoother, which also leads to a loss of details. In this case, we choose the restored image with α=1.3 as the best image since it provides the highest PSNR value. Throughout the experiments, we select α=1.3 or 1.5. Finally, λ3 and λ4 control the separation of stripes from the image. For λ3, we test four different values, such as 0.05,0.1,0.6, 1. If the value of λ3 is too small, the stripes are not extracted properly and the restored image is over-smoothed. But a large enough value of λ3 enables a successful extraction of stripes. Indeed, for λ3=0.6 or higher, the denoising results do not change much, so we fix the value of λ3 to 0.6 throughout the experiment. For λ4, we test four values, such as 0.01, 0.05, 0.2, 0.4. It can be observed that using λ1=0.01 or 0.4 fails to properly extract stripes from the image, while using λ4=0.05 or 0.2 provides better decomposition of the stripe and image components than the others. Besides, λ4=0.05 and 0.2 provide very similar PSNR values. Although the parameter λ4 is more sensitive than the other parameters, λ4 is selected from {0.05,0.08,0.2,0.4} in many cases.
This section presents the denoising results in the presence of non-periodic stripe noise and Gaussian noise.
First, Figure 11 presents the denoising results tested on the vatican image in the presence of non-periodic stripes with r=50 or 70, m=100, and Gaussian noise with σ=10. The noisy data images are provided in the first row. We can see that all the models except our model fail to properly separate the stripes from the image, which leads to some traces of stripes in the restored images. This is also visible in the difference images between u and u∗. For all cases, our model effectively eliminates both stripes and Gaussian noise, resulting in the highest PSNR and SSIM values. These show the efficiency of our directional term and nonconvex group sparsity term of stripes to extract non-periodic stripes in the presence of Gaussian noise.
In Figure 12, we present the denoising results tested on two other different images in the presence of non-periodic stripes with r=50 or 70, m=50, and Gaussian noise with σ=20. Similarly, TVGS eliminates stripes from the images better than LRSID, LRGS+TV, Schatten+TV, and ELRTV+TV, bringing better denoised images. But there are traces of steaks in both the restored and difference images of TVGS. In contrast, our model removes both stripe noise and Gaussian noise sufficiently, yielding cleaner restored images than other models. Furthermore, our model mitigates the staircase artifacts that appeared in the restored images of TVGS. Therefore, these examples also confirm the effectiveness of the proposed model for removing both non-periodic stripes and Gaussian noise.
Figure 13 shows the column mean cross-track profiles of the restored images (blue curve) in Figure 12 and original images (red curve). The horizontal axis represents the column number, and the vertical axis represents the mean value of the intensities in each column. It can be seen that the curves of our model are similar to the original ones. Meanwhile, there are large gaps between the curves of the other models and the original ones. Hence, these examples also show better denoising performance of our model than the others.
In Table 3, we record the average of the PSNR and SSIM values of all methods tested on all images in Figure 2, in the presence of non-periodic stripes and Gaussin noise. We can see that the proposed model supplies the highest PSNR and SSIM values for all cases. The quantitative assessment is also consistent with visual results. This illustrates that the proposed model is superior to the existing models in terms of visual quality and image quality evaluation.
r | 30 | 50 | 70 | ||||
m | 50 | 100 | 50 | 100 | 50 | 100 | |
σ=10 | LRSID | 25.17/0.7666 | 23.60/0.7457 | 24.76/0.7612 | 22.86/0.7367 | 24.38/0.7590 | 21.94/0.7245 |
TVGS | 28.93/0.8831 | 28.87/0.8827 | 28.45/0.8782 | 28.23/0.8764 | 27.50/0.8704 | 25.96/0.8613 | |
LRGS | 26.93/0.8077 | 26.74/0.8065 | 26.63/0.8058 | 26.08/0.8005 | 25.95/0.8010 | 24.51/0.7914 | |
LRGS+TV | 27.08/0.8598 | 26.89/0.8586 | 26.78/0.8578 | 26.21/0.8490 | 26.07/0.8514 | 24.59/0.8371 | |
Shatten | 26.51/0.7940 | 25.97/0.7919 | 26.02/0.7904 | 25.57/0.7824 | 26.13/0.7911 | 24.14/0.7403 | |
Shatten+TV | 27.28/0.8674 | 26.69/0.8643 | 26.63/0.8576 | 26.33/0.8561 | 26.83/0.8634 | 25.06/0.8218 | |
ELRTV | 25.59/0.7918 | 24.90/0.8375 | 25.34/0.7902 | 24.55/0.7822 | 25.11/0.7880 | 21.73/0.7775 | |
ELRTV+TV | 25.79/0.8450 | 25.08/0.8375 | 25.54/0.8433 | 24.69/0.8305 | 25.29/0.8405 | 23.83/0.8245 | |
Our | 29.19/0.8867 | 29.19/0.8869 | 29.01/0.8843 | 28.89/0.8828 | 28.40/0.8788 | 28.54/0.8798 | |
σ=20 | LRSID | 24.57/0.7557 | 23.11/0.7343 | 24.13/0.7498 | 22.49/0.7246 | 24.05/0.7495 | 21.52/0.7146 |
TVGS | 25.53/0.7770 | 25.45/0.7763 | 25.05/0.7692 | 25.06/0.7683 | 24.58/0.7613 | 24.15/0.7484 | |
LRGS | 22.33/0.6149 | 22.31/0.6147 | 22.29/0.6148 | 22.29/0.6146 | 22.24/0.6145 | 22.23/0.6139 | |
LRGS+TV | 24.83/0.7625 | 24.81/0.7630 | 24.75/0.7619 | 24.76/0.7619 | 24.67/0.7607 | 24.66/0.7608 | |
Shatten | 21.73/0.6000 | 21.73/0.6001 | 21.41/0.5992 | 21.11/0.5885 | 21.58/0.5994 | 20.61/0.5624 | |
Shatten+TV | 24.52/0.7545 | 24.49/0.7547 | 24.01/0.7525 | 23.78/0.7453 | 24.24/0.7528 | 23.55/0.7350 | |
ELRTV | 21.92/0.6064 | 21.81/0.6047 | 21.93/0.6066 | 21.91/0.6060 | 21.93/0.6066 | 21.90/0.6062 | |
ELRTV+TV | 24.29/0.7572 | 24.11/0.7551 | 24.30/0.7573 | 24.27/0.7567 | 24.30/0.7569 | 24.27/0.7570 | |
Our | 25.75/0.7851 | 25.74/0.7851 | 25.29/0.7771 | 25.46/0.7794 | 25.05/0.7717 | 25.08/0.7633 |
Table 4 presents the computing time of all models, in the case of the non-periodic stripe noise with (r,m)=(70,50) and Gaussian noise with σ=20. It can be observed that LRSID, TVGS, LRGS(+TV), and ELRTV(+TV) models are faster than Schatten(+TV) and our model. Despite the high computational cost, the proposed model provides better restoration results than other models.
Image | LRSID | TVGS | LRGS (+TV) | Shatten (+TV) | ELRTV (+TV) | Our |
MODISBAND20 | 17.36 | 20.24 | 24.37 (25.06) | 24.62 (25.47) | 17.51 (18.31) | 26.96 |
BAND20 | 15.54 | 19.34 | 20.40 (20.94) | 28.48 (28.95) | 17.97 (18.59) | 26.95 |
Original band30 | 7.17 | 7.05 | 10.11 (10.40) | 10.54 (10.82) | 8.78 (9.15) | 9.54 |
rio | 15.36 | 19.25 | 21.12 (21.67) | 29.06 (29.51) | 17.35 (17.98) | 26.79 |
rio1 | 15.47 | 17.20 | 25.48 (26.18) | 30.18 (30.88) | 19.55 (20.18) | 28.14 |
helliniko | 16.17 | 19.45 | 22.15 (22.68) | 32.24 (32.87) | 17.98 (18.56) | 25.83 |
helliniko1 | 16.82 | 19.29 | 20.70 (21.26) | 23.74 (24.17) | 17.60 (18.17) | 26.54 |
vatican | 16.64 | 20.21 | 19.69 (20.21) | 29.50 (30.14) | 18.14 (18.72) | 27.13 |
image02162021 | 19.38 | 23.84 | 25.81 (26.73) | 32.28 (33.29) | 20.18 (21.26) | 32.89 |
image02172021 | 16.85 | 17.56 | 17.11 (17.71) | 30.54 (31.25) | 16.85 (17.55) | 26.65 |
Figure 14 presents the plots of the PSNR and energy functional values of our model via the outer iteration number. As the outer iteration number increases, we can see that the PSNR values gradually increase and converge to some constant values, while the energy values gradually decrease. These plots justify the numerical convergence of the proposed algorithm.
In Figure 15, we present the extracted stripe noise components, s, of all models from Figures 6, 7, 11, and 12. We also record the PSNR and SSIM values between the extracted stripe noise and the originally added stripe noise. It can be observed that the extracted stripes of our model are very close to the originally added stripes, which contributes the highest PSNR and SSIM values of our stripe noise component. This confirms the effectiveness of our model for extracting stripe noise in the presence of Gaussian noise.
Finally, in Figure 16, we present the denoising results for color SAR images in the presence of non-periodic stripes with (r,m)=(70,100) and Gaussian noise with σ=20. Stripes and Gaussian noise are added to each color channel independently, and all models are applied to each channel. We can see that our model not only appropriately separates the stripes from the images but also preserves edges and details well compared with other models. These examples also validate the effectiveness of the proposed model on color SAR images with a mixture of stripes and Gaussian noise. More experimental results on color SAR images can be found in the material at the following link: https://url.kr/kdgrcz.
In this paper, we introduce an image decomposition model to extract stripe noise and image components from a noisy image that is corrupted by a mixture of stripe noise and Gaussian noise. We considered various types of periodic or non-periodic stripe noise and also assumed a relatively high level of Gaussian noise unlike the previous works. For the stripe noise component, a unidirectional TV and a nonconvex group sparsity term were exploited, and they enabled the proper separation of periodic or non-periodic stripes from images with high levels of Gaussian noise. Furthermore, for the image component, we made use of a nonconvex FTV regularization, which not only ameliorated the staircase effect appearing in images recovered from TV-based models but also enabled the conservation of edges and details. To handle the nonconvex and nonsmooth problem, we adopted IRL1 and ADMM. This led to an efficient iterative algorithm capable of satisfactorily solving the proposed model, and we also proved its global convergence. The numerical results validated that the proposed model generated superior denoising results than other existing models in terms of visual and image quality assessments. Despite the effective performance of the proposed model, issues remain regarding high computational time and many parameters, which need to be investigated in future work.
All authors developed the theoretical formalism, performed the analytic calculations and performed the numerical simulations. All authors contributed to the final version of the manuscript.
The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.
Myeongmin Kang was supported by the National Research Foundation (NRF) of Korea grant (No. 2019R1I1A3A01055168). Miyoun Jung was supported by the Hankuk University of Foreign Studies Research Fund and the National Research Foundation (NRF) of Korea grant (No. RS-2023-00241770).
The authors declare no conflicts of interest in this paper.
[1] | J. A. Richards, Remote sensing digital image analysis: an introduction, Springer Berlin, Heidelberg, 2013, 343–380. https://doi.org/10.1007/978-3-642-30062-2 |
[2] |
I. Makki, R. Younes, C. Francis, T. Bianchi, M. Zucchetti, A survey of landmine detection using hyperspectral imaging, ISPRS J. Photogramm. Remote Sens., 124 (2017), 40–53. https://doi.org/10.1016/j.isprsjprs.2016.12.009 doi: 10.1016/j.isprsjprs.2016.12.009
![]() |
[3] |
H. Zhang, J. Li, Y. Huang, L. Zhang, A nonlocal weighted joint sparse representation classification method for hyperspectral imagery, IEEE J-STARS, 7 (2017), 2056–2065. https://doi.org/10.1109/JSTARS.2013.2264720 doi: 10.1109/JSTARS.2013.2264720
![]() |
[4] |
Y. Tarabalka, J. Chanussot, J. A. Benediktssons, Segmentation and classification of hyperspectral images using watershed transformation, Pattern Recogn., 43 (2010), 2367–2379. https://doi.org/10.1016/j.patcog.2010.01.016 doi: 10.1016/j.patcog.2010.01.016
![]() |
[5] |
D. W. J. Stein, S. G. Beaven, L. E. Hoff, E. M. Winter, A. P. Schaum, A. D. Stocker, Anomaly detection from hyperspectral imagery, IEEE Signal Proc. Mag., 19 (2002), 58–69. https://doi.org/10.1109/79.974730 doi: 10.1109/79.974730
![]() |
[6] |
M. D. Iordache, J. M. Bioucas-Dias, A. Plaza, Collaborative sparse regression for hyperspectral unmixing, IEEE Trans. Geosci. Remote Sens., 52 (2014), 341–354. https://doi.org/10.1109/TGRS.2013.2240001 doi: 10.1109/TGRS.2013.2240001
![]() |
[7] |
J. Chen, Y. Shao, H. Guo, W. Wang, B. Zhu, Destriping CMODIS data by power filtering, IEEE Trans. Geosci. Remote Sens., 41 (2003), 2119–2124. https://doi.org/10.1109/TGRS.2003.817206 doi: 10.1109/TGRS.2003.817206
![]() |
[8] |
J. Chen, H. Lin, Y. Shao, L. Yang, Oblique striping removal in remote sensing imagery based on wavelet transform, Int. J. Remote Sens., 27 (2006), 1717–1723. https://doi.org/10.1080/01431160500185516 doi: 10.1080/01431160500185516
![]() |
[9] |
R. Pande-Chhetri, A. Abd-Elrahman, De-striping hyperspectral imagery using wavelet transform and adaptive frequency domain filtering, ISPRS J. Photogramm. Remote Sens., 66 (2011), 620–636. https://doi.org/10.1016/j.isprsjprs.2011.04.003 doi: 10.1016/j.isprsjprs.2011.04.003
![]() |
[10] |
L. Sun, R. Neville, K. Staenz, H. P. White, Automatic destriping of Hyperion imagery based on spectral moment matching, J. Can. Remote Sens., 34 (2008), S68–S81. https://doi.org/10.5589/m07-067 doi: 10.5589/m07-067
![]() |
[11] |
M. Wegener, Destriping multiple sensor imagery by improved histogram matching, Int. J. Remote Sens., 11 (1990), 859–875. https://doi.org/10.1080/01431169008955060 doi: 10.1080/01431169008955060
![]() |
[12] |
H. Shen, L. Zhang, A MAP-based algorithm for destriping and inpainting of remotely sensed images, IEEE Trans. Geosci. Remote Sens., 47 (2009), 1492–1502. https://doi.org/10.1109/TGRS.2008.2005780 doi: 10.1109/TGRS.2008.2005780
![]() |
[13] |
M. Bouali, S. Ladjal, Toward optimal destriping of MODIS data using a unidirectional variational model, IEEE Trans. Geosci. Remote Sens., 49 (2011), 2924–2935. https://doi.org/10.1109/TGRS.2011.2119399 doi: 10.1109/TGRS.2011.2119399
![]() |
[14] |
Y. Chang, H. Fang, L. Yan, H. Liu, Robust destriping method with unidirectional total variation and framelet regularization, Opt. Express, 21 (2013), 23307–23323. https://doi.org/10.1364/OE.21.023307 doi: 10.1364/OE.21.023307
![]() |
[15] |
Y. Chang, L. Yan, H. Fang, H. Liu, Simultaneous destriping and denoising for remote sensing images with unidirectional total variation and sparse representation, IEEE Geosci. Remote Sens. Lett., 11 (2014), 1051–1055. https://doi.org/10.1109/LGRS.2013.2285124 doi: 10.1109/LGRS.2013.2285124
![]() |
[16] |
Y. Zhang, G. Zhou, L. Yan, T. Zhang, A destriping algorithm based on TV-Stokes and unidirectional total variation model, Optik, 127 (2016), 428–439. https://doi.org/10.1016/j.ijleo.2015.09.246 doi: 10.1016/j.ijleo.2015.09.246
![]() |
[17] |
M. Wang, X. Zheng, J. Pan, B. Wang, Unidirectional total variation destriping using difference curvature in MODIS emissive bands, Infrared Phys. Technol., 75 (2016), 1–11. https://doi.org/10.1016/j.infrared.2015.12.004 doi: 10.1016/j.infrared.2015.12.004
![]() |
[18] |
X. Liu, X. Lu, H. Shen, Q. Yuan, Y. Jiao, L. Zhang, Stripe noise separation and removal in remote sensing images by consideration of the global sparsity and local variational properties, IEEE Trans. Geosci. Remote Sens., 54 (2016), 3049-3060. https://doi.org/10.1109/TGRS.2015.2510418 doi: 10.1109/TGRS.2015.2510418
![]() |
[19] |
Y. Chang, L. Yan, T. Wu, S. Zhong, Remote sensing image stripe noise removal: from image decomposition perspective, IEEE Trans. Geosci. Remote Sens., 54 (2016), 7018–7031. https://doi.org/10.1109/TGRS.2016.2594080 doi: 10.1109/TGRS.2016.2594080
![]() |
[20] |
Y. Chen, T. Z. Huang, X. Zhao, L. J. Deng, J. Huang, Stripe noise removal of remote sensing images by total variation regularization and group sparsity constraint, Remote sens., 9 (2017), 559. https://doi.org/10.3390/rs9060559 doi: 10.3390/rs9060559
![]() |
[21] |
Y. Chen, T. Z. Huang, L. J. Deng, X. L. Zhao, M. Wang, Group sparsity based regularization model for remote sensing image stripe noise removal, Neurocomputing, 267 (2017), 95–106. https://doi.org/10.1016/j.neucom.2017.05.018 doi: 10.1016/j.neucom.2017.05.018
![]() |
[22] |
Y. Chen, T. Z. Huang, X. L. Zhao, Destriping of multispectral remote sensing image using low-rank tensor decomposition, IEEE J-STARS, 11 (2018), 4950–4967. https://doi.org/10.1109/JSTARS.2018.2877722 doi: 10.1109/JSTARS.2018.2877722
![]() |
[23] |
H. X. Dou, T. Z. Huang, L. J. Deng, X. L. Zhao, J. Huang, Directional ℓ0 sparse modeling for image stripe noise removal, Remote Sens., 10 (2018), 361. https://doi.org/10.3390/rs10030361 doi: 10.3390/rs10030361
![]() |
[24] |
S. Qiong, Y. Wang, X. Yan, H. Gu, Remote sensing images stripe noise removal by double sparse regulation and region separation, Remote Sens., 10 (2018), 998. https://doi.org/10.3390/rs10070998 doi: 10.3390/rs10070998
![]() |
[25] |
J. Wang, T. Z. Huang, T. H. Ma, X. L. Zhao, Y. Chen, A sheared low-rank model for oblique stripe removal, Appl. Math. Comput., 360 (2019), 167–180. https://doi.org/10.1016/j.amc.2019.03.066 doi: 10.1016/j.amc.2019.03.066
![]() |
[26] |
J. L. Wang, T. Z. Huang, X. L. Zhao, J. Huang, T. H. Ma, Y. B. Zheng, Reweighted block sparsity regularization for remote sensing images destriping, IEEE J-STARS, 12 (2019), 4951–4963. https://doi.org/10.1109/JSTARS.2019.2940065 doi: 10.1109/JSTARS.2019.2940065
![]() |
[27] |
J. H. Yang, X. L. Zhao, T. H. Ma, Y. Chen, T. Z. Huang, M. Ding, Remote sensing images destriping using unidirectional hybrid total variation and nonconvex low-rank regularization, J. Comput. Appl. Math., 363 (2020), 124–144. https://doi.org/10.1016/j.cam.2019.06.004 doi: 10.1016/j.cam.2019.06.004
![]() |
[28] |
X. Wu, H. Qu, L. Zheng, T. Gao, A remote sensing image destriping model based on low-rank and directional sparse constraint, Remote Sens., 13 (2021), 5126. https://doi.org/10.3390/rs13245126 doi: 10.3390/rs13245126
![]() |
[29] | X. Liu, X. Lu, H. Shen, Q. Yuan, L. Zhang, Oblique stripe removal in remote sensing images via oriented variation, arXiv, 2018. https://doi.org/10.48550/arXiv.1809.02043 |
[30] |
Q. Song, Z. Huang, H. Ni, K. Bai, Z. Li, Remote sensing images destriping with an enhanced low-rank prior and total variation regulation, Signal Image Video Process., 16 (2022), 1895–1903. https://doi.org/10.1007/s11760-022-02149-8 doi: 10.1007/s11760-022-02149-8
![]() |
[31] |
L. Song, H. Huang, Simultaneous destriping and image denoising using a nonparametric model with the EM algorithm, IEEE Trans. Image Process., 32 (2023), 1065–1077. https://doi.org/10.1109/TIP.2023.3239193 doi: 10.1109/TIP.2023.3239193
![]() |
[32] |
N. Kim, S. S. Han, C. S. Jeong, ADOM: ADMM-Based optimization model for stripe noise removal in remote sensing image, IEEE Access, 11 (2023), 106587–106606. https://doi.org/10.1109/ACCESS.2023.3319268 doi: 10.1109/ACCESS.2023.3319268
![]() |
[33] |
F. Yan, S. Wu, Q. Zhang, Y. Liu, H. Sun, Destriping of remote sensing images by an optimized variational model, Sensors, 23 (2023), 7529. https://doi.org/10.3390/s23177529 doi: 10.3390/s23177529
![]() |
[34] | C. Wang, X. Zhao, Q. Wang, Z. Ma, P. Tang, An inexact proximal majorization-minimization algorithm for remote sensing image stripe noise removal, Numer. Algor., 2024. https://doi.org/10.1007/s11075-023-01743-2 |
[35] |
L. I. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algorithm, Phys. D, 60 (1992), 259–268. https://doi.org/10.1016/0167-2789(92)90242-F doi: 10.1016/0167-2789(92)90242-F
![]() |
[36] |
T. Chan, A. Marquina, P. Mulet, High-order total variation-based image restoration, SIAM J. Sci. Comput., 22 (2000), 503–516. https://doi.org/10.1137/S1064827598344169 doi: 10.1137/S1064827598344169
![]() |
[37] |
M. Lysaker, A. Lundervold, X. C. Tai, Noise removal using fourth-order partial differential equation with application to medical magnetic resonance images in space and time, IEEE Trans. Image Process., 12 (2003), 1579–1590. https://doi.org/10.1109/TIP.2003.819229 doi: 10.1109/TIP.2003.819229
![]() |
[38] |
K. Bredies, K. Kunisch, T. Pock, Total generalized variation, SIAM J. Imaging Sci., 3 (2010), 492–526. https://doi.org/10.1137/090769521 doi: 10.1137/090769521
![]() |
[39] |
F. Li, C. Shen, J. Fan, C. Shen, Image restoration combining a total variational filter and a fourth-order filter, J. Vis. Commun. Image Represent., 18 (2007), 322–330. https://doi.org/10.1016/j.jvcir.2007.04.005 doi: 10.1016/j.jvcir.2007.04.005
![]() |
[40] |
K. Papafitsoros, C. B. Sch¨onlieb, A combined first and second order variational approach for image restoration, J. Math. Imaging Vis., 48 (2014), 308–338. https://doi.org/10.1007/s10851-013-0445-4 doi: 10.1007/s10851-013-0445-4
![]() |
[41] |
J. Bai, X. C. Feng, Fractional-order anisotropic diffusion for image denoising, IEEE Trans. Image Process., 16 (2007), 2492–2502. https://doi.org/10.1109/TIP.2007.904971 doi: 10.1109/TIP.2007.904971
![]() |
[42] |
J. Zhang, Z. Wei, L. Xiao, Adaptive fractional-order multi-scale method for image denoising, SIAM J. Imaging Sci., 43 (2012), 39–49. https://doi.org/10.1007/s10851-011-0285-z doi: 10.1007/s10851-011-0285-z
![]() |
[43] |
R. H. Chan, A. Lanza, S. Morigi, F. Sgallari, An adaptive strategy for the restoration of textured images using fractional order regularization, Numer. Math. Theor. Meth. Appl., 6 (2013), 276–296. https://doi.org/10.4208/nmtma.2013.mssvm15 doi: 10.4208/nmtma.2013.mssvm15
![]() |
[44] |
J. Zhang, Z. Hui, L. Xiao, A fast adaptive reweighted residual-feedback iterative algorithm for fractional order total variation regularized multiplicative noise removal of partly-textured images, Signal Process., 98 (2014), 381–395. https://doi.org/10.1016/j.sigpro.2013.12.009 doi: 10.1016/j.sigpro.2013.12.009
![]() |
[45] |
J. Zhang, K. Chen, A total fractional-order variation model for image restoration with nonhomogeneous boundary conditions and its numerical solution, SIAM J. Imaging Sci., 8 (2015), 2487–2518. https://doi.org/10.1137/14097121X doi: 10.1137/14097121X
![]() |
[46] |
A. Ullah, W. Chen, M. A. Khan, A new variational approach for restoring images with multiplicative noise, Comput. Math. Appl., 71 (2016), 2034–2050. https://doi.org/10.1016/j.camwa.2016.03.024 doi: 10.1016/j.camwa.2016.03.024
![]() |
[47] |
F. Dong, Y. Chen, A fractional-order derivative based variational framework for image denoising, Inverse Probl. Imag., 10 (2016), 27–50. https://doi.org/10.3934/ipi.2016.10.27 doi: 10.3934/ipi.2016.10.27
![]() |
[48] |
M. R. Chowdhury, J. Zhang, J. Qin, Y. Lou, Poisson image denoising based on fractional-order total variation, Inverse Probl. Imag., 14 (2020), 77–96. https://doi.org/10.3934/ipi.2019064 doi: 10.3934/ipi.2019064
![]() |
[49] |
Y. F. Pu, J. L. Zhou, X. Yuan, Fractional differential mask: a fractional differential-based approach for multiscale texture enhancement, IEEE Trans. Image Process., 19 (2010), 491–511. https://doi.org/10.1109/TIP.2009.2035980 doi: 10.1109/TIP.2009.2035980
![]() |
[50] |
Z. Ren, C. He, Q. Zhang, Fractional order total variation regularization for image super-resolution, Signal Process., 93 (2013), 2408–2421. https://doi.org/10.1016/j.sigpro.2013.02.015 doi: 10.1016/j.sigpro.2013.02.015
![]() |
[51] |
D. Geman, G. Reynolds, Constrained restoration and recovery of discontinuities, IEEE Trans. Pattern Anal. Mach. Intell., 14 (1992), 367–383. https://doi.org/10.1109/34.120331 doi: 10.1109/34.120331
![]() |
[52] |
D. Geman, C. Yang, Nonlinear image recovery with half-quadratic regularization, IEEE Trans. Image Process., 4 (1995), 932–946. https://doi.org/10.1109/83.392335 doi: 10.1109/83.392335
![]() |
[53] |
M. Nikolova, M. K. Ng, C. P. Tam, Fast nonconvex nonsmooth minimization methods for image restoration and reconstruction, IEEE Trans. Image Process., 19 (2010), 3073–3088. https://doi.org/10.1109/TIP.2010.2052275 doi: 10.1109/TIP.2010.2052275
![]() |
[54] |
S. Oh, H. Woo, S. Yun, M. Kang, Non-convex hybrid total variation for image denoising, J. Vis. Commun. Image Represent., 24 (2013), 332–344. https://doi.org/10.1016/j.jvcir.2013.01.010 doi: 10.1016/j.jvcir.2013.01.010
![]() |
[55] |
M. Kang, M. Kang, M. Jung, Nonconvex higher-order regularization based Rician noise removal with spatially adaptive parameters, J. Visual Commun. Image Represent., 32 (2015), 180–193. https://doi.org/10.1016/j.jvcir.2015.08.006 doi: 10.1016/j.jvcir.2015.08.006
![]() |
[56] |
T. Adam, R. Paramesran, Hybrid non-convex second-order total variation with applications to non-blind image deblurring, Signal Image Video Process., 14 (2020), 115–123. https://doi.org/10.1007/s11760-019-01531-3 doi: 10.1007/s11760-019-01531-3
![]() |
[57] |
Y. Sun, L. Lei, D. Guan, X. Li, G. Xiao, SAR image speckle reduction based on nonconvex hybrid total variation model, IEEE Trans. Geosci. Remote Sens., 59 (2020), 1231–1249. https://doi.org/10.1109/TGRS.2020.3002561 doi: 10.1109/TGRS.2020.3002561
![]() |
[58] |
P. Ochs, A. Dosovitskiy, T. Brox, T. Pock, On iteratively reweighted algorithms for nonsmooth nonconvex optimization in computer vision, SIAM J. Imaging Sci., 8 (2015), 331–372. https://doi.org/10.1137/140971518 doi: 10.1137/140971518
![]() |
[59] |
E. J. Candés, M. B. Wakin, S. P. Boyd, Enhancing sparsity by reweighted ℓ1 minimization, J. Fourier Anal. Appl., 14 (2008), 877–905. https://doi.org/10.1007/s00041-008-9045-x doi: 10.1007/s00041-008-9045-x
![]() |
[60] |
J. Eckstein, D. P. Bertsekas, On the Douglas-Rachford splitting method and the proximal point algorithm for maximal monotone operators, Math. Program., 55 (1992), 293–318. https://doi.org/10.1007/BF01581204 doi: 10.1007/BF01581204
![]() |
[61] | R. Glowinski, Numerical methods for nonlinear variational problems, Springer Berlin, Heidelberg, 1984. https://doi.org/10.1007/978-3-662-12613-4 |
[62] |
S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein, Distributed optimization and statistical learning via the alternating direction method of multipliers, Found. Trends Mach. Learn., 3 (2011), 1–122. https://doi.org/10.1561/2200000016 doi: 10.1561/2200000016
![]() |
[63] |
H. Carfantan, J. Idier, Statistical linear destriping of satellite-based pushbroom-type images, IEEE Trans. Geosci. Remote Sens., 48 (2010), 1860–1871. https://doi.org/10.1109/TGRS.2009.2033587 doi: 10.1109/TGRS.2009.2033587
![]() |
[64] | K. Miller, B. Ross, An introduction to the fractional calculus and fractional differential equations, New York, USA: John Wiley & Sons, 1993. |
[65] | K. B. Oldham, J. Spanier, The fractional calculus: theory and applications of differentiation and integration to arbitrary order, New York, USA: Academic Press, 1974. https://doi.org/10.1016/s0076-5392(09)x6012-1 |
[66] | I. Podlubny, Fractional differential equations: an introduction to fractional derivatives, fractional differential equations, to methods of their solution and some of their applications, London, UK: Academic Press, 1999. https://doi.org/10.1016/s0076-5392(99)x8001-5 |
[67] | L. Vese, T. F. Chan, Reduced non-convex functional approximations for image restoration & segmentation, UCLA CAM Report, 1997. |
[68] |
H. Attouch, J. Bolte, B. F. Svaiter, Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward–backward splitting, and regularized Gauss–Seidel methods, Math. Program., 137 (2013), 91–129. https://doi.org/10.1007/s10107-011-0484-9 doi: 10.1007/s10107-011-0484-9
![]() |
[69] | L. P. D. Van den Dries, Tame topology and o-minimal structures, New York, NY, USA: Cambridge University Press, 1998. https://doi.org/10.1017/CBO9780511525919 |
[70] |
A. Chambolle, T. Pock, A first-order primal-dual algorithm for convex problems with applications to imaging, J. Math. Imaging Vis., 40 (2011), 120–145. https://doi.org/10.1007/s10851-010-0251-1 doi: 10.1007/s10851-010-0251-1
![]() |
[71] |
T. Goldstein, S. Osher, The split Bregman method for L1-regularized problems, SIAM J. Imaging Sci., 2 (2009), 323–343. https://doi.org/10.1137/080725891 doi: 10.1137/080725891
![]() |
[72] |
Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli, Image quality assessment: from error visibility to structural similarity, IEEE Trans. Image Process., 13 (2004), 600–612. https://doi.org/10.1109/TIP.2003.819861 doi: 10.1109/TIP.2003.819861
![]() |
1. | Yating Zhu, Zixun Zeng, Zhong Chen, Deqiang Zhou, Jian Zou, Performance analysis of the convex non-convex total variation denoising model, 2024, 9, 2473-6988, 29031, 10.3934/math.20241409 |
r | 30 | 50 | 70 | ||||
m | 50 | 100 | 50 | 100 | 50 | 100 | |
σ=10 | LRSID | 28.20/0.8760 | 27.27/0.8640 | 27.86/0.8733 | 27.35/0.8665 | 28.02/0.8749 | 27.41/0.8664 |
TVGS | 29.06/0.8849 | 29.02/0.8848 | 28.78/0.8809 | 28.74/0.8807 | 28.59/0.8784 | 28.51/0.8775 | |
LRGS | 26.94/0.8075 | 26.93/0.8073 | 26.86/0.8070 | 26.84/0.8071 | 26.72/0.8062 | 26.70/0.8060 | |
LRGS+TV | 27.09/0.8599 | 27.09/0.8600 | 27.01/0.8588 | 26.98/0.8586 | 26.85/0.8575 | 26.83/0.8574 | |
Shatten | 26.69/0.7945 | 26.31/0.7919 | 26.43/0.7920 | 26.59/0.7943 | 26.56/0.7941 | 26.44/0.7923 | |
Shatten+TV | 27.51/0.8688 | 27.04/0.8656 | 27.09/0.8594 | 27.37/0.8678 | 27.34/0.7941 | 27.11/0.7923 | |
ELRTV | 25.73/0.7918 | 25.51/0.7896 | 25.77/0.7923 | 25.74/0.7921 | 25.78/0.7923 | 25.75/0.7921 | |
ELRTV+TV | 25.94/0.8456 | 25.71/0.8433 | 25.99/0.8458 | 25.96/0.8455 | 25.99/0.8459 | 25.97/0.8458 | |
Our | 29.26/0.8878 | 29.26/ 0.8879 | 29.06/0.8847 | 29.07/0.8849 | 28.90/ 0.8825 | 28.87/0.8822 | |
σ=20 | LRSID | 25.01/0.7595 | 24.43/0.7480 | 24.92/0.7595 | 24.50/0.7525 | 25.05/0.7614 | 24.65/0.7552 |
TVGS | 25.60/0.7783 | 25.57/0.7784 | 25.42/0.7742 | 25.38/0.7732 | 25.28/0.7702 | 25.21/0.7695 | |
LRGS | 22.33/0.6149 | 22.31/0.6147 | 22.29/0.6148 | 22.29/0.6146 | 22.24/0.6145 | 22.23/0.6139 | |
LRGS+TV | 24.83/0.7625 | 24.81/0.7630 | 24.75/0.7619 | 24.76/0.7619 | 24.67/0.7607 | 24.66/0.7608 | |
Shatten | 21.86/0.6022 | 21.73/0.6001 | 21.84/0.6024 | 21.80/0.6010 | 21.83/0.6016 | 21.81/0.6015 | |
Shatten+TV | 24.72/0.7564 | 24.49/0.7547 | 24.67/0.7566 | 24.63/0.7566 | 24.71/0.7570 | 24.64/0.7566 | |
ELRTV | 21.92/0.6064 | 21.81/0.6047 | 21.93/0.6066 | 21.91/0.6060 | 21.93/0.6066 | 21.90/0.6062 | |
ELRTV+TV | 24.29/0.7572 | 24.11/0.7551 | 24.30/0.7573 | 24.27/0.7567 | 24.30/0.7569 | 24.27/0.7570 | |
Our | 25.79/0.7861 | 25.78/0.7864 | 25.63/0.7820 | 25.62/0.7817 | 25.41/0.7766 | 25.45/0.7775 |
Image | P | LRSID | TVGS | LRGS+TV | Shatten+TV | ELRTV+TV | Our |
(a) | 16 | 24.80/0.6388 | 26.72/0.6921 | 25.69/0.6921 | 24.81/0.6635 | 24.81/0.6774 | 27.15/0.7072 |
32 | 25.38/0.6660 | 26.43/0.6919 | 25.99/0.6966 | 25.69/0.6843 | 25.69/0.6984 | 27.07/0.7088 | |
64 | 23.54/0.6157 | 26.32/0.6628 | 25.81/0.6876 | 25.60/0.6801 | 25.57/0.6905 | 27.04/0.7066 | |
(b) | 16 | 25.35/0.6999 | 26.45/0.6963 | 26.34/0.7034 | 25.08/0.6797 | 25.09/0.7084 | 26.77/0.7032 |
32 | 26.23/0.7042 | 26.41/0.6944 | 26.64/0.7038 | 25.83/0.6826 | 26.04/0.7109 | 26.70/0.7028 | |
64 | 23.10/0.6851 | 26.00/0.6926 | 26.32/0.7031 | 25.95/0.6834 | 25.76/0.7096 | 26.70/0.7029 | |
(g) | 16 | 23.91/0.7935 | 25.05/0.8025 | 24.67/0.8044 | 24.04/0.7968 | 24.09/0.8066 | 25.24/0.8095 |
32 | 24.00/0.7805 | 25.01/0.8014 | 24.93/0.7780 | 24.69/0.7972 | 24.81/0.8069 | 25.20/0.8076 | |
64 | 21.40/0.7714 | 24.92/0.8016 | 24.85/0.8045 | 24.70/0.7982 | 24.77/0.8082 | 25.21/0.8092 |
r | 30 | 50 | 70 | ||||
m | 50 | 100 | 50 | 100 | 50 | 100 | |
σ=10 | LRSID | 25.17/0.7666 | 23.60/0.7457 | 24.76/0.7612 | 22.86/0.7367 | 24.38/0.7590 | 21.94/0.7245 |
TVGS | 28.93/0.8831 | 28.87/0.8827 | 28.45/0.8782 | 28.23/0.8764 | 27.50/0.8704 | 25.96/0.8613 | |
LRGS | 26.93/0.8077 | 26.74/0.8065 | 26.63/0.8058 | 26.08/0.8005 | 25.95/0.8010 | 24.51/0.7914 | |
LRGS+TV | 27.08/0.8598 | 26.89/0.8586 | 26.78/0.8578 | 26.21/0.8490 | 26.07/0.8514 | 24.59/0.8371 | |
Shatten | 26.51/0.7940 | 25.97/0.7919 | 26.02/0.7904 | 25.57/0.7824 | 26.13/0.7911 | 24.14/0.7403 | |
Shatten+TV | 27.28/0.8674 | 26.69/0.8643 | 26.63/0.8576 | 26.33/0.8561 | 26.83/0.8634 | 25.06/0.8218 | |
ELRTV | 25.59/0.7918 | 24.90/0.8375 | 25.34/0.7902 | 24.55/0.7822 | 25.11/0.7880 | 21.73/0.7775 | |
ELRTV+TV | 25.79/0.8450 | 25.08/0.8375 | 25.54/0.8433 | 24.69/0.8305 | 25.29/0.8405 | 23.83/0.8245 | |
Our | 29.19/0.8867 | 29.19/0.8869 | 29.01/0.8843 | 28.89/0.8828 | 28.40/0.8788 | 28.54/0.8798 | |
σ=20 | LRSID | 24.57/0.7557 | 23.11/0.7343 | 24.13/0.7498 | 22.49/0.7246 | 24.05/0.7495 | 21.52/0.7146 |
TVGS | 25.53/0.7770 | 25.45/0.7763 | 25.05/0.7692 | 25.06/0.7683 | 24.58/0.7613 | 24.15/0.7484 | |
LRGS | 22.33/0.6149 | 22.31/0.6147 | 22.29/0.6148 | 22.29/0.6146 | 22.24/0.6145 | 22.23/0.6139 | |
LRGS+TV | 24.83/0.7625 | 24.81/0.7630 | 24.75/0.7619 | 24.76/0.7619 | 24.67/0.7607 | 24.66/0.7608 | |
Shatten | 21.73/0.6000 | 21.73/0.6001 | 21.41/0.5992 | 21.11/0.5885 | 21.58/0.5994 | 20.61/0.5624 | |
Shatten+TV | 24.52/0.7545 | 24.49/0.7547 | 24.01/0.7525 | 23.78/0.7453 | 24.24/0.7528 | 23.55/0.7350 | |
ELRTV | 21.92/0.6064 | 21.81/0.6047 | 21.93/0.6066 | 21.91/0.6060 | 21.93/0.6066 | 21.90/0.6062 | |
ELRTV+TV | 24.29/0.7572 | 24.11/0.7551 | 24.30/0.7573 | 24.27/0.7567 | 24.30/0.7569 | 24.27/0.7570 | |
Our | 25.75/0.7851 | 25.74/0.7851 | 25.29/0.7771 | 25.46/0.7794 | 25.05/0.7717 | 25.08/0.7633 |
Image | LRSID | TVGS | LRGS (+TV) | Shatten (+TV) | ELRTV (+TV) | Our |
MODISBAND20 | 17.36 | 20.24 | 24.37 (25.06) | 24.62 (25.47) | 17.51 (18.31) | 26.96 |
BAND20 | 15.54 | 19.34 | 20.40 (20.94) | 28.48 (28.95) | 17.97 (18.59) | 26.95 |
Original band30 | 7.17 | 7.05 | 10.11 (10.40) | 10.54 (10.82) | 8.78 (9.15) | 9.54 |
rio | 15.36 | 19.25 | 21.12 (21.67) | 29.06 (29.51) | 17.35 (17.98) | 26.79 |
rio1 | 15.47 | 17.20 | 25.48 (26.18) | 30.18 (30.88) | 19.55 (20.18) | 28.14 |
helliniko | 16.17 | 19.45 | 22.15 (22.68) | 32.24 (32.87) | 17.98 (18.56) | 25.83 |
helliniko1 | 16.82 | 19.29 | 20.70 (21.26) | 23.74 (24.17) | 17.60 (18.17) | 26.54 |
vatican | 16.64 | 20.21 | 19.69 (20.21) | 29.50 (30.14) | 18.14 (18.72) | 27.13 |
image02162021 | 19.38 | 23.84 | 25.81 (26.73) | 32.28 (33.29) | 20.18 (21.26) | 32.89 |
image02172021 | 16.85 | 17.56 | 17.11 (17.71) | 30.54 (31.25) | 16.85 (17.55) | 26.65 |
r | 30 | 50 | 70 | ||||
m | 50 | 100 | 50 | 100 | 50 | 100 | |
σ=10 | LRSID | 28.20/0.8760 | 27.27/0.8640 | 27.86/0.8733 | 27.35/0.8665 | 28.02/0.8749 | 27.41/0.8664 |
TVGS | 29.06/0.8849 | 29.02/0.8848 | 28.78/0.8809 | 28.74/0.8807 | 28.59/0.8784 | 28.51/0.8775 | |
LRGS | 26.94/0.8075 | 26.93/0.8073 | 26.86/0.8070 | 26.84/0.8071 | 26.72/0.8062 | 26.70/0.8060 | |
LRGS+TV | 27.09/0.8599 | 27.09/0.8600 | 27.01/0.8588 | 26.98/0.8586 | 26.85/0.8575 | 26.83/0.8574 | |
Shatten | 26.69/0.7945 | 26.31/0.7919 | 26.43/0.7920 | 26.59/0.7943 | 26.56/0.7941 | 26.44/0.7923 | |
Shatten+TV | 27.51/0.8688 | 27.04/0.8656 | 27.09/0.8594 | 27.37/0.8678 | 27.34/0.7941 | 27.11/0.7923 | |
ELRTV | 25.73/0.7918 | 25.51/0.7896 | 25.77/0.7923 | 25.74/0.7921 | 25.78/0.7923 | 25.75/0.7921 | |
ELRTV+TV | 25.94/0.8456 | 25.71/0.8433 | 25.99/0.8458 | 25.96/0.8455 | 25.99/0.8459 | 25.97/0.8458 | |
Our | 29.26/0.8878 | 29.26/ 0.8879 | 29.06/0.8847 | 29.07/0.8849 | 28.90/ 0.8825 | 28.87/0.8822 | |
σ=20 | LRSID | 25.01/0.7595 | 24.43/0.7480 | 24.92/0.7595 | 24.50/0.7525 | 25.05/0.7614 | 24.65/0.7552 |
TVGS | 25.60/0.7783 | 25.57/0.7784 | 25.42/0.7742 | 25.38/0.7732 | 25.28/0.7702 | 25.21/0.7695 | |
LRGS | 22.33/0.6149 | 22.31/0.6147 | 22.29/0.6148 | 22.29/0.6146 | 22.24/0.6145 | 22.23/0.6139 | |
LRGS+TV | 24.83/0.7625 | 24.81/0.7630 | 24.75/0.7619 | 24.76/0.7619 | 24.67/0.7607 | 24.66/0.7608 | |
Shatten | 21.86/0.6022 | 21.73/0.6001 | 21.84/0.6024 | 21.80/0.6010 | 21.83/0.6016 | 21.81/0.6015 | |
Shatten+TV | 24.72/0.7564 | 24.49/0.7547 | 24.67/0.7566 | 24.63/0.7566 | 24.71/0.7570 | 24.64/0.7566 | |
ELRTV | 21.92/0.6064 | 21.81/0.6047 | 21.93/0.6066 | 21.91/0.6060 | 21.93/0.6066 | 21.90/0.6062 | |
ELRTV+TV | 24.29/0.7572 | 24.11/0.7551 | 24.30/0.7573 | 24.27/0.7567 | 24.30/0.7569 | 24.27/0.7570 | |
Our | 25.79/0.7861 | 25.78/0.7864 | 25.63/0.7820 | 25.62/0.7817 | 25.41/0.7766 | 25.45/0.7775 |
Image | P | LRSID | TVGS | LRGS+TV | Shatten+TV | ELRTV+TV | Our |
(a) | 16 | 24.80/0.6388 | 26.72/0.6921 | 25.69/0.6921 | 24.81/0.6635 | 24.81/0.6774 | 27.15/0.7072 |
32 | 25.38/0.6660 | 26.43/0.6919 | 25.99/0.6966 | 25.69/0.6843 | 25.69/0.6984 | 27.07/0.7088 | |
64 | 23.54/0.6157 | 26.32/0.6628 | 25.81/0.6876 | 25.60/0.6801 | 25.57/0.6905 | 27.04/0.7066 | |
(b) | 16 | 25.35/0.6999 | 26.45/0.6963 | 26.34/0.7034 | 25.08/0.6797 | 25.09/0.7084 | 26.77/0.7032 |
32 | 26.23/0.7042 | 26.41/0.6944 | 26.64/0.7038 | 25.83/0.6826 | 26.04/0.7109 | 26.70/0.7028 | |
64 | 23.10/0.6851 | 26.00/0.6926 | 26.32/0.7031 | 25.95/0.6834 | 25.76/0.7096 | 26.70/0.7029 | |
(g) | 16 | 23.91/0.7935 | 25.05/0.8025 | 24.67/0.8044 | 24.04/0.7968 | 24.09/0.8066 | 25.24/0.8095 |
32 | 24.00/0.7805 | 25.01/0.8014 | 24.93/0.7780 | 24.69/0.7972 | 24.81/0.8069 | 25.20/0.8076 | |
64 | 21.40/0.7714 | 24.92/0.8016 | 24.85/0.8045 | 24.70/0.7982 | 24.77/0.8082 | 25.21/0.8092 |
r | 30 | 50 | 70 | ||||
m | 50 | 100 | 50 | 100 | 50 | 100 | |
σ=10 | LRSID | 25.17/0.7666 | 23.60/0.7457 | 24.76/0.7612 | 22.86/0.7367 | 24.38/0.7590 | 21.94/0.7245 |
TVGS | 28.93/0.8831 | 28.87/0.8827 | 28.45/0.8782 | 28.23/0.8764 | 27.50/0.8704 | 25.96/0.8613 | |
LRGS | 26.93/0.8077 | 26.74/0.8065 | 26.63/0.8058 | 26.08/0.8005 | 25.95/0.8010 | 24.51/0.7914 | |
LRGS+TV | 27.08/0.8598 | 26.89/0.8586 | 26.78/0.8578 | 26.21/0.8490 | 26.07/0.8514 | 24.59/0.8371 | |
Shatten | 26.51/0.7940 | 25.97/0.7919 | 26.02/0.7904 | 25.57/0.7824 | 26.13/0.7911 | 24.14/0.7403 | |
Shatten+TV | 27.28/0.8674 | 26.69/0.8643 | 26.63/0.8576 | 26.33/0.8561 | 26.83/0.8634 | 25.06/0.8218 | |
ELRTV | 25.59/0.7918 | 24.90/0.8375 | 25.34/0.7902 | 24.55/0.7822 | 25.11/0.7880 | 21.73/0.7775 | |
ELRTV+TV | 25.79/0.8450 | 25.08/0.8375 | 25.54/0.8433 | 24.69/0.8305 | 25.29/0.8405 | 23.83/0.8245 | |
Our | 29.19/0.8867 | 29.19/0.8869 | 29.01/0.8843 | 28.89/0.8828 | 28.40/0.8788 | 28.54/0.8798 | |
σ=20 | LRSID | 24.57/0.7557 | 23.11/0.7343 | 24.13/0.7498 | 22.49/0.7246 | 24.05/0.7495 | 21.52/0.7146 |
TVGS | 25.53/0.7770 | 25.45/0.7763 | 25.05/0.7692 | 25.06/0.7683 | 24.58/0.7613 | 24.15/0.7484 | |
LRGS | 22.33/0.6149 | 22.31/0.6147 | 22.29/0.6148 | 22.29/0.6146 | 22.24/0.6145 | 22.23/0.6139 | |
LRGS+TV | 24.83/0.7625 | 24.81/0.7630 | 24.75/0.7619 | 24.76/0.7619 | 24.67/0.7607 | 24.66/0.7608 | |
Shatten | 21.73/0.6000 | 21.73/0.6001 | 21.41/0.5992 | 21.11/0.5885 | 21.58/0.5994 | 20.61/0.5624 | |
Shatten+TV | 24.52/0.7545 | 24.49/0.7547 | 24.01/0.7525 | 23.78/0.7453 | 24.24/0.7528 | 23.55/0.7350 | |
ELRTV | 21.92/0.6064 | 21.81/0.6047 | 21.93/0.6066 | 21.91/0.6060 | 21.93/0.6066 | 21.90/0.6062 | |
ELRTV+TV | 24.29/0.7572 | 24.11/0.7551 | 24.30/0.7573 | 24.27/0.7567 | 24.30/0.7569 | 24.27/0.7570 | |
Our | 25.75/0.7851 | 25.74/0.7851 | 25.29/0.7771 | 25.46/0.7794 | 25.05/0.7717 | 25.08/0.7633 |
Image | LRSID | TVGS | LRGS (+TV) | Shatten (+TV) | ELRTV (+TV) | Our |
MODISBAND20 | 17.36 | 20.24 | 24.37 (25.06) | 24.62 (25.47) | 17.51 (18.31) | 26.96 |
BAND20 | 15.54 | 19.34 | 20.40 (20.94) | 28.48 (28.95) | 17.97 (18.59) | 26.95 |
Original band30 | 7.17 | 7.05 | 10.11 (10.40) | 10.54 (10.82) | 8.78 (9.15) | 9.54 |
rio | 15.36 | 19.25 | 21.12 (21.67) | 29.06 (29.51) | 17.35 (17.98) | 26.79 |
rio1 | 15.47 | 17.20 | 25.48 (26.18) | 30.18 (30.88) | 19.55 (20.18) | 28.14 |
helliniko | 16.17 | 19.45 | 22.15 (22.68) | 32.24 (32.87) | 17.98 (18.56) | 25.83 |
helliniko1 | 16.82 | 19.29 | 20.70 (21.26) | 23.74 (24.17) | 17.60 (18.17) | 26.54 |
vatican | 16.64 | 20.21 | 19.69 (20.21) | 29.50 (30.14) | 18.14 (18.72) | 27.13 |
image02162021 | 19.38 | 23.84 | 25.81 (26.73) | 32.28 (33.29) | 20.18 (21.26) | 32.89 |
image02172021 | 16.85 | 17.56 | 17.11 (17.71) | 30.54 (31.25) | 16.85 (17.55) | 26.65 |