Moreau-Enhanced Total Variation and Subspace Factorization for Hyperspectral Denoising

: Hyperspectral images (HSIs) denoising aims at recovering noise-free images from noisy counterparts to improve image visualization. Recently, various prior knowledge has attracted much attention in HSI denoising, e.g., total variation (TV), low-rank, sparse representation, and so on. However, the computational cost of most existing algorithms increases exponentially with increasing spectral bands. In this paper, we fully take advantage of the global spectral correlation of HSI and design a uniﬁed framework named subspace-based Moreau-enhanced total variation and sparse factorization (SMTVSF) for multispectral image denoising. Speciﬁcally, SMTVSF decomposes an HSI image into the product of a projection matrix and abundance maps, followed by a ‘Moreau-enhanced’ TV (MTV) denoising step, i.e., a nonconvex regularizer involving the Moreau envelope mechnisam, to reconstruct all the abundance maps. Furthermore, the schemes of subspace representation penalizing the low-rank characteristic and (cid:96) 2,1 -norm modelling the structured sparse noise are embedded into our denoising framework to reﬁne the abundance maps and projection matrix. We use the augmented Lagrange multiplier (ALM) algorithm to solve the resulting optimization problem. Extensive results under various noise levels of simulated and real hypspectral images demonstrate our superiority against other competing HSI recovery approaches in terms of quality metrics and visual effects. In addition, our method has a huge advantage in computational efﬁciency over many competitors, beneﬁting from its removal of most spectral dimensions during iterations.


Introduction
With the wealth of spatial and spectral information, hyperspectral image (HSI) delivers a more accurate description ability of real scenes to distinguish precise details than color images and provides potential advantages of application in vegetation monitoring, medical diagnosis, mineral exploration, among numerous others [1,2]. However, due to the instrument restrictions and various weather conditions, HSIs are unavoidably corrupted by complicated forms of noise during the acquisition procedure [3], which significantly plagues the subsequent applications, such as classification [4], spectral unmixing [5,6], and anomaly detection [7,8]. As a result, denoising the HSIs become an essential preprocessing step to support the subsequent HSI-related applications and analysis.
There is ever-growing attention in HSI denoising and numerous approaches have been developed in recent decade. A straightforward approach is to use a classical gray-level image restoration method for HSI in a bandwise manner. See, e.g., [9][10][11]. However, this strategy always leads to unsatisfactory results as the correlations among adjacent bands or pixels have been ignored. Aiming at improving the recovery quality, scholars turn to simultaneously encode the spatial and spectral correlation knowledge property enhancement. He et al. [42] transfer the non-local denoising to the projected HSI instead of the original spectral dimension to reduce processing cost. However, the methods mentioned above cannot handle structural sparse noise. Therefore, the denoising performance based on subspace representation still has room for improvement.
Motivated by such studies, a non-separable non-convex TV denoising penalty combined with the subspace representation framework is proposed to recover HSI from mixed noise. We first decompose the original HSI as a subspace matrix multiplied by coefficients matrix and subsequently explore the spatial piecewise smoothness effectively in the projected space. Specifically, the non-separable non-convex TV penalty is generated based on the Moreau envelope, which follows a similar idea proposed in our previous paper [43]. Our main contributions are as follows: 1.
The low-rank structure of HSI data are explored via subspace representation, which is realized by decomposing the HSI into the product of projection matrix and abundance maps. To separate the noise out, the projection matrix is required to be orthogonal.

2.
The Moreau-enhanced TV as a non-convex regularizer is applied to each abundance map, which is expected to enhance the spatial information and thus achieve more performance boost. Furthermore, subspace representation and 2,1 -norm are incorporated into our denoising framework.

3.
A variational augmented Lagrangian method (ALM) algorithm is designed to minimize the proposed model. Extensive experiments confirm that our method can offer an excellent recovery performance compared with many popular HSI denoising methods. More importantly, the proposed algorithm requires less computational cost than most others.
This paper is organized as follows. Section 2 reviews the HSI degradation model and some related works. Section 3 presents the proposed Moreau-enhanced total variation and subspace factorization model, followed by its optimization solution. Various experimental results including comparisons with other competing methods are reported in Section 4. The final section gives our conclusions.

Problem Formulation
Let Y ∈ R p×h×w represent an observed HSI data with p spectral bands of spatial size h × w. In this study, the 3D HSI image is rearranged as a 2D matrix Y ∈ R p×l (l = h × w) for the purpose of easier analysis, where each row contains a vectorized spectral channel band. Following the studies of [31,37], HSIs are assumed to be contaminated by mixed noise, and its noisy observation can be described in a matrix form as where X represents the clean HSI; S denotes sparse error term, representing a mixture of impulse noise and outliers; and N is the Gaussian noise. The images X, S, and N are modeled the same way as Y.
The goal of HSI denoising is to separate the noise from observed image with proper constraints. Under the assumption of image degradation model (1), the generalization model for estimating the clean image X is designed as where the first term is the data-fidelity term, which represents the similarity between the restored image and degraded image, and · F denotes the Frobenius norm. J 1 (X) is a regularization term describing the intrinsic structure of the image component, e.g., non-local similarity [44], global low-rank property [45]. J 2 (S) is the sparse regularization characterizing the prior knowledge of the sparse corruptions. λ 1 and λ 2 are two predefined scalar parameters used to balance the contribution of each relevant term. Therefore, the core of image denoising is how to exactly exploit the prior information for both X and S to effectively separate the noise components.

Global Low-Rank Matrix Factorization
HSIs are well-known for holding strong spectral low-rank characteristic. This suggests that each spectral signature (columns of X) lies in a common k-dimensional subspace S k . As stated in [46,47], the value of k is far less than the spectral dimension, i.e., p k. Therefore, the target image X can be presented as X = AM, where A ∈ R p×k represents the basis of the subspace S k , hereafter termed projection matrix, and M ∈ R k×l is regarded as the coefficient matrix. In many real situations, both A and M have evident physical meanings. Specifically, the columns of A are the spectral signatures (endmembers), and M is the abundance fraction vector. According to the idea presented in [40], the projection matrix A is required to be orthogonal, i.e., A T A = I, and it can be obtained by singular value decomposition (SVD) or other signal subspace identification approaches [48] of hyperspectral imagery.
As a consequence, the HSI data can be modeled as Following [42], the orthogonal constraint A T A = I can keep noise out of A and help to further refine the abundance map iteratively. By incorporating the orthogonal constraint into the model (2), the corresponding HSI restoration model is denoted as where J 1 (M) describes the prior information of the abundance map and · 2,1 refers to the 2,1 -norm, which is associated with the sparse noise of the spatial difference images. It should be also pointed out that model (4) includes capturing the global correlation along spectrum as well as reducing the computational burden.

Moreau-Enhanced TV Restoration Model
TV-based regularization is well known to enforce piecewise smoothness and has been successfully used in HSI processing [31,37]. Denoting the image domain as y ∈ R h × w , the general image restoration model of linear total variation function is formulated as tvd(y; η) = arg min which claims that TV restoration model can be regarded as the proximity operator of x → x TV . η is a regularization parameter that maintains the trade-off between the Frobenius norm and the TV norm.
To precisely preserve the local piecewise continuity and induce sparsity more efficiently than the TV norm, we define a non-differentiable non-convex penalty suitable for TV denoising model, i.e., where S α (x) indicates the Moreau envelop of · TV with index 1/α as the threshold (α > 0), and is defined as Proposition 1. Let α ≥0, the function S α (x) is convex and differentiable. Its gradient ∇S α (x) can be expressed as Proof. From the Proposition 12.15 of [49], S α (x) is convex, real-valued, continuous, and exact since S α (x) is the Moreau envelop of · TV with index 1/α. Similarly, from the Proposition 12.29 of [49], For α = 0, S 0 (x) = 0. Following with the definition of gradient, we have ∇S 0 (x) = 0. Thus, we obtain the proposition.
Simultaneously, if α > 0, Ψ α (x) is non-convex, and the 'degree of concavity' of the functional Ψ α (x) is controlled by the value of scalar coefficient α. When α = 0, model (6) reverts to the classical TV penalty. Compared with TV norm, if α > 0, the new penalty can more accurately determine the jumps of the piecewise constant image.
To strongly induce sparsity, we introduce the new penalty function Ψ α (x) as the piecewise smoothness surrogate to substitute the TV norm in (5). Consequently, the Moreau-enhanced TV denoising model is formulated as mtvd(y; η, α) = arg min where the parameter 0 ≤α ≤ 1/η is such that the model (10) is convex; 0 ≤ α < 1/η is such that the model (10) is strictly convex. For the non-convex analysis of model (10), we refer the readers to [50] and the references therein. When model (10) is strongly convex, the minimizer problem can be efficiently solved by the forward-backward splitting (FBS) algorithm [49]. Its solution is given in Appendix A.

Subspace-Based Moreau-Enhanced TV Restoration Model
According to the description of Section 2.2, matrix M ∈ R k×l can be inferred from the projection of the observed HSI Y onto subspace A, i.e., M = A T Y. In addition, the dimension k of subspace A can be estimated by HySime [48] (We give the reference of the HySime algorithm). The details can be found in Appendix B. According to [51], the abundance map inherits two important properties of the original image, namely sparsity and piecewise smooth structure. To guarantee lower complexity and better performance, in this paper, 'Moreau-enhanced' TV denoising technique is applied on matrix M to preserve the above-mentioned two characteristics concurrently.
After defining the 'Moreau-enhanced' TV denoising model in Section 3.1, we treat it as a prior and process abundance maps in a band-wise manner. Mathematically, abundance maps denoising can be reformed asL in which M i represents the row vector of the abundance matrix, F : R l → R h×w denotes the operation reshaping a vector to a 2D matrix with a size of h×w, and L represents the denoised result. To verify the superiority of 'Moreau-enhanced' TV denoising model (Equation (10)) over TV denoising model (Equation (5)), we utilize the two models to restore the abundance matrix M and projection A (shown in Figure 1 Simplifying model (11) yields the following representation: With all these, the proposed SMTVSF HSI denoising model can be written as follows: where λ 1 controls the contribution of 'Moreau-enhanced' TV denoising model, and λ 2 controls the weight of the sparsity term. Theoretically, the second term of model (13) will degenerate into the band-by-band TV denoising model when α is set to 0. Our final model unifies matrix factorization, Gaussian noise removal, and sparse noise elimination in a framework; thus, it is expected to have a strong ability to separate mixed noise from the degraded image. The basic idea of the model is to first transform the observed HSI into certain low-dimensional space, and subsequently recover the spatial detail information of the projected HSI. To enhance the performance, we refine the abundance maps and projection space during the iteration. Then, the restored HSI is obtained via inverse transformation. Figure 2 illustrates the overview of the proposed method.

Optimization Procedure
The proposed method is mainly composed of three components: spectral low-rank projection; spatial domain denoising; and sparsity constraint of the sparsity noise. The details of the first step can be found in Section 2.2. The solution of the other two steps is presented in the following.
(1) Spatial denoising via abundance maps: In this stage, the 'Moreau-enhanced' TV regularized method is applied to the abundance map instead of the whole spectral image to reduce the calculation cost. The objective in this stage iŝ The parameter α affects the denoising performance and we analyze the influence of parameter α in Section 4.3. It should be noted that Ψ α (x) degrades into the standard TV penalty when α is set to 0. Then, the above model (14) can be converted into k standard 2D TV denoising problems: For solving model (16), we introduce the ALM methodology [52] to minimize the following model: Subsequently, we split the multivariable optimization problem (17) into several manageable subproblems and the solution of the corresponding subfunctions can be established in an alternating minimization manner.
(1) The update of S is given bŷ Let ς = Y − AM, following [53], the optimum solution of S-subproblem can be applied independently to each column and its j-th columnŜ(:, j) can be solved bŷ (2) Similarly, the optimal solution of M is given bŷ whose solution can be formulated aŝ (3) As to the subproblem A, we havê Here, the classic singular value decomposition (SVD) is employed to The pseudocode for solving the proposed SMTVSF model is summarized in the following (Algorithm 1).  (21) 8: Update A as Equation (22) 9: Obtain the denoised HSI X = AM.

Experiment Results
We conduct several groups of experiments on both simulated and real images to demonstrate the performance of the proposed method compared with some existing typical algorithms. The following representative HSI denoising methods are selected as competitors, i.e., low-rank matrix recovery (LRMR) [18], TV-regularized low-rank matrix factorization (LRTV) [31], fast hyperspectral image denoising based on low-rank and sparse representations (FastHyDe) [46], global matrix factorization and to local tensor factorizations method (GLF) [41], spatial-spectral total variation regularized local low-rank matrix recovery (LLRGTV) [37], Moreau-enhanced total variation regularized local low-rank matrix recovery (LLRMTV) [43], and non-local meets global (NM-meet) denoising paradigm [42]. LRMR is one of the representative methods to recover HSI by the global low rank and sparse optimization. LRTV belongs to a band manner TV regularized low-rank restoration method. FastHyDe and GLF exploit low-rank and sparse representations to restore HSIs. FastHyDe refines representation coefficients with nonlocal patch-based methods, while GLF employs low-rank tensor factorization of non-local similar patches to fully induce the self-similar of HSI. LLRGTV and LLRMTV divide the HSI into local overlapping patches to separate the low-rank clean signal from the sparse noise. LLRGTV unites the SSTV-regularized method to update all patches simultaneously, while LLRMTV reconstructs the whole HSI by a Moreau-enhanced total variation denoising method. NM-meet combines global spectral low-rank projection and WNNM [34] to denoise the noisy HSI. We have implemented LLRMTV and obtained the other codes from the author's homepages. The parameters of all competitors are set by using default values or fine-tuning to report the optimal results, while, for our SMTVSF solver, we will discuss the parameter sensitivity to the performance in Section 4.3. The experiments are programmed in MATLAB R2016b (MathWorks, Natick, Massachusetts State, US) on a laptop with 8 GB RAM, Intel Core i5 processor @2.9 GHz (Santa Clara, California, US). Before restoration, the whole HSIs are linearly mapped into [0, 1].

Simulated Experiments
(1) Data Description: Two clean HSI datasets are selected as the ground-truth images, i.e., Washington DC Mall data set [54] and Pavia city center data set [55] The original Washington DC Mall image contains 1208 × 307 pixels in the spatial domain and 191 bands in the spectral domain. Following [37], a patch of size 256 × 256 × 191 is considered in the experiment. Following the ideas of [37,46], the first 22 bands (containing all the noisy bands) of the Pavia city center dataset image are manually removed before denoising, which ensures that the remaining spectral bands can be used as the reference. Then, we use a subimage of size 200 × 200 × 80 to perform the comparison.
Four types of synthetic noisy observation are used to facilitate the overall evaluation of all methods, both in visual quality and quantitative evaluation. A detailed description is elaborated as follows: Case 1: Each spectral band of the test image is corrupted by Zero-mean Gaussian noise with the variance being 0.05.
Case 2: Zero-mean Gaussian noise with variance value being 0.1 is added to all bands. In addition, all bands are further contaminated by the impulse noise with a consistent ratio of 0.05.
Case 3: Gaussian noise with different noise levels and impulse noise with different ratios are added to the HSI data. The variance of Gaussian noise and the ratio of impulse noise varies randomly from 0 to 0.2.
Case 4: Mixtures of Gaussian with impulse noises are added just as in Case 3. In addition, we randomly chose 20 bands to add deadlines with a width of 1-3.
(2) Visual comparison: To compare the overall performance, two representative cases 3 and 4 are selected for visual inspection. It can be evident that the original images suffer great contaminations in both cases as presented in Figures 3b-6b. In addition, the magnified subregion of each subfigure is given for a more detailed comparison. Figures 3 and 5 display the performance of all competitors acting on images with mixtures of Gaussian and impulse noises. From Figure 3, we can observe that all competing methods can remove the great mass of noise. Among them, LRMR and FastHyDe come up several image distortion and over-smoothness. In addition, in comparison with the original image, LRTV, GLF, and NM-meet both smooth the image details to an extent. However, taking Figure 5g as an instance, it is obvious that some noise still remains in the restored image. In the meantime, different degrees of blurring can still be observed in the demarcated windows of Figure 5c-f. SMTVSF leads to a more faithful recovery (Figures 3 and 5j) of the ground truth (Figures 3 and 5a). Moreover, LLRGTV and LLRMTV can also successfully remove this mixed noise from the damaged band and preserve details as we expected. This further demonstrates the advantage of the TV-regularization in noise removal.    Figure 4, it is easy to observe that FastHyDe, GLF, and NM-meet cannot completely remove this complex mixture noise, which contain some evident deadlines and introduce some stripe-like artifacts as shown in Figure 4e-g. LRMR obtains better results than the above three methods, but cannot thoroughly remove the deadlines. LRTV yields minor distortion compared with LRMR but blurs many details compared with the original clean image. In the meantime, a similar observation also appears in Figure 6. For LLRMTV, the deadlines are removed completely, but some residual noise is still nonnegligible, as can be seen in the zoom-in view of Figure 4i. In comparison, the proposed SMTVSF obtains the best visual quality and retains the essential edge structures. In addition, the LLRGTV method also has comparable performance, which is consistent with the following quantitative comparison. (3) Quantitative comparison: Three objective quality metrics, i.e., peak-signal-to-noise ratio (PSNR), structural similarity index measurement (SSIM), and erreur relative globale adimensionnelle de synthese (ERGAS) are used in our study to evaluate the restoration accuracy. The smaller ERGAS, the larger mean PSNR (MPSNR), and mean SSIM (MSSIM), the closer the reference one.
For each noise setting of both simulated datasets, Table 1 lists the assessment results by the competing denoising methods, with the best results labeled in bold. These assessment values have great consistency with the visual effects, which manifest the superiority of the proposed method once again. In general, NM-meet shows worse denoising performance under complex mixed noise. Another observation is that FastHyDe and GLF can achieve comparable better performance in pure Gaussian noise cases, while, with the intensity of stripes and impulse noises increasing, the inferiority of the above-mentioned three methods is evident. LLRMTV and LLRGTV produce similar measurements. However, LLRMTV takes much more runtime since it introduces a new nonconvex regularizer. Meanwhile, we can observe that the proposed method is more competitive than LRMR, LRTV, and LLRGTV in complex mixed noise. This phenomenon validates the advantage of both subspace representation and nonconvex regularization for HSI restoration. Interestingly, the efficiency of our approach is only second to FastHyDe, but it outperforms all the competitors in qualitative metrics. In addition, we also report the corresponding quality metrics of the STVSF model, which can be directly considered as a combination of standard TV constraint and sparse representation. As seen, the performance of STVSF is relatively lower than SMTVSF. It once again verifies the effectiveness of the nonconvex regularization. (4) Qualitative Comparison: Furthermore, the PSNR and SSIM measurements across all bands with different restoration algorithms of both datasets on case 4 are plotted in Figure 7. Note that GLF performs slightly better than FastHyDe. For simplicity, we do not show the visual restoration results of FastHyDe. From Figure 7a,b, it can be seen that SMTVSF obtains the higher PSNR and SSIM values on almost all bands of the Washington DC Mall data. For Pavia city center data, we obtain an average optimum in all channels as shown in Figure 7c,d.

Real Experiments
Here, we chose two real noisy HSI data to test the practicability of the proposed method, i.e., the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) Indian Pines dataset and the Hyperspectral Digital Imagery Collection Experiment (HYDICE) urban dataset. Similar to the synthetic experiments, we use default parameters or fine-tuned parameters to implement other competing methods, and the whole HSIs are also scaled to the interval [0, 1]. Since the subspace dimension of the real dataset needs to be predefined, the HySime algorithm [48] is adopted to estimate it.
(1) AVIRIS Indian Pines Dataset: We apply the proposed method to the AVIRIS Indian Pines data set [56]. This is an image with a size of 145 × 145 × 220 and some bands are severely contaminated by Gaussian noise, salt and pepper, and unknown noise. The subspace dimension of this image is set to 25, which is generated by the HySime algorithm.
A visual comparison of bands 3 and 219 for all competing methods are given in Figures 8  and 9, respectively. From Figure 8a, we can see that this band is degraded by Gaussian noise and sparse corruptions. LRMR and LRTV perform unsatisfactorily in the restoration. FastHyDe and GLF can obviously improve the spatial resolution, but some edges are not clearly differentiated, as the red dashed boxes shown in Figure 8d,e. NM-meet can moderately preserve the details, but when image is heavily contaminated by mixed noise, it loses its practicality, as presented in Figure 9f. The underlying reason is that NGmeet aims to eliminate Gaussian noise. Obviously, when the noise level increases, LRMR, LRTV, FastHyDe, and GLF result in a less competitive performance (see Figures 8 and 9b-e). The results of LLRMTV, LLRGTV, and SMTVSF have relatively less noise. We can also see that the restored HSI of LLRGTV maintains some horizontal stripes, as presented in Figure 8g. Meanwhile, SMTVSF introduces some artifacts, as presented in Figure 9i. Comparatively, for this data set, LLRMTV can wipe out the noise efficiently and deliver a better visual appearance.
(2) HYDICE Urban Dataset: The urban area HYDICE dataset [57] is provided by Chen et al. [21] and seriously polluted by the atmosphere, stripes and dead lines. The size of the scene is 307 × 307 × 210, and we feed a subimage of 200 × 200 × 210 into the denoising algorithms. In this set of experiments, the value of the subspace dimension is set to 4.
We also select two typical bands to depict the denoising performance of different methods. Figure 10 shows the band 139 before and after denoising, while Figure 11 provides a visual comparison of the band 208. It is not hard to see that LRMR, GLF, and NM-meet cannot effectively remove the stripes, and also introduce artifacts into the final results. LLRGTV can remove random noise, but it still shows stripes remaining as presented in Figures 10g and 11g. This is because the stripes and deadlines in this scene exist in the same location in some adjacent bands and the low-rank based method treats the structured sparse noise as a clean image component. LRTV and FastHyDe perform comparatively better in destriping. However, LRTV would over-smooth the details. LLRMTV can remove this mixed noise and the corresponding edges are relatively clear as shown in Figures 10h and 11h. Benefitting from the 'Moreau-enhanced' Total Variation and sparse representation regularizations, the proposed SMTVSF method can eliminate complicated noise and finely recover the details hiding under the polluted HSI.   Figure 12 shows a more intuitive verification with respect to the horizontal mean profiles of the band 139 before and after applying all of the compared techniques. The evident fluctuations in Figure 12a indicate that severely stripes are located in the image. From the restored curves, it can be observed that LRMR, LRTV, GLF, and NM-meet fail to keep a right curve tendency. By contrast, FastHyDe generates noticeable distortion. LLRMTV can obtain a relatively smoother curve than SMTVSF, which agrees with the visual results drawn in Figure 10. Our comprehensive experiments show that SMTVSF can achieve some competitive performance compared with other popular methods (e.g., LLRGTV), which indicates the effectiveness of SMTVSF for HSI restoration. (3) Running Time: Table 2 compares the processing time of all methods on two real-world HSIs. From the table, we can conclude that FastHyDe is the fastest method and our method is slightly slower than it. However, we provide better restored results than FastHyDe. This is mainly because our algorithm requires the use of a non-separable non-convex TV denoising penalty to converge to the final solution.

Parameter Sensitivity
Several critical parameters in the restoration model (13) need to be carefully identified, including regularization parameters λ 1 and λ 2 , subspace dimension k, and non-convex control parameter α. In the following, we first present the optimal values of these parameter settings and then display the convergence curve of our method. Please note that all the results are based on simulated Washington DC Mall case 4.
(1) Parameter λ 1 and λ 2 : We tune λ 1 from 0.05 to 0.5 with an interval of 0.05 and λ 2 from 0.1 to 0.3 with an interval of 0.025, respectively. Figure 13 shows the curves of the MPSNR and MSSIM values with changing λ 1 and λ 2 . Clearly, the proposed method performs appealing results when  (2) Subspace Dimension k: Subspace dimension k has an inherent relationship with the size of the projection matrix. A detailed construction of the subspace A is listed in Appendix B. Figure 14 presents the influence of parameters k from 1 to 30 on the denoising performance in terms of MPSNR and MSSIM values. As can be observed, the two objective metrics achieve the optimal performance when k is set to 5. However, this is different from the evaluation result k = 4 of the HySime algorithm [48], mainly because it only considers Gaussian noise. In our experiments, the subspace dimension needs to be assigned manually according to the data itself. (3) Parameter α: α is the non-convex control parameter, which directly influences the solution of (15). Empirically fixing η as 0.125 × √ MN, we vary α within the scope of (0, 1/η) and show the performance in Figure 15. Intuitively, when α falls within the range of [0.4, 0.95]/η, the MPSNR values of the proposed method are relatively stable. When α increases, the time consumption decreases sharply as model (11) will be strongly convex and have a unique minimizer (see Figure 15b). Taking both denoising performance and computational complexity into consideration, α is set to 0.7/η in all our experiments. (4) Convergence Analysis: In Figure 16, we display the changes of the objective function and MPSNR values with the iteration number. Here, we can easily observe that, with the increment of iteration number, these values increase rapidly. After the number is greater than 20, the performance is relatively stable, indicating that our method can be successfully converged.

Conclusions
In this paper, we propose a subspace-based method, namely SMTVSF, for HSI denoising. Our SMTVSF model integrates the spectral low-rank property and spatial sparsity into a general fomulation. Specifically, these two terms are explored by subspace representation and representation coefficients denoising, respectively. Our technological process starts by decomposing an HSI into two submatrices with lower dimensions, thereby reducing the computational burden. Subsequently, for each subspace representation image, a Moreau-enhanced Total Variation denoising model is utilized to promote the spatial piecewise smoothness and consistency of the HSI. In addition, 2,1 -norm regularization is further embedded to distinguish the stripe component. Extensive experiments verify that, compared with other competitors, our method can both achieve a better accuracy performance and obtain an acceptable running time. Inspired the spatial-spectral TV regularization idea, in our future work, we plan to develop a 3D Moreau-enhanced TV denoising model used in HSI denoising for deeper feature learning both spatially and spectrally.
FBS is guaranteed to converge for 0 < µ < 2/ρ and ρ is the Lipschitz constant of ∇f 1 . We implement the algorithm with µ = 1 because this value yields faster convergence. In practice, the iterates x (k) often converge quickly to the optimal solution of the objective function.