betaFeedback

WorksEngineering

Deep Plug-and-Play Proximal Distance Algorithm for Robust Principal Component Analysis

Abstract

Robust Principal Component Analysis (RPCA) is essential in signal processing, with applications in hyperspectral imaging, face recognition, and video background modeling. Traditional RPCA methods often suffer from insufficient utilization of data priors and struggle to capture complex image structures. Conversely, advanced deep denoisers like DnCNN, FFDNet, and DRUNet excel at leveraging image data priors but lack physical model constraints and demonstrate poor robustness against structured noise in RPCA tasks. This paper proposes a novel Plug-and-Play (PnP) method for multi-channel RPCA, integrating both the physical generative model and cutting-edge deep priors. Extensive experiments on RGB and hyperspectral datasets show that PnP-PDA achieves better quantitative performance than conventional RPCA methods and state-of-the-art methods, including deep denoisers and tensor RPCA approaches, while preserving visual details.

Introduction

ROBUST principal component analysis (RPCA) is a fundamental topic in signal processing, with numerous applications in hyperspectral imaging [1], robust face recognition [2], and video background modeling [3]. In RPCA, an image is typically decoupled into two components: a low-rank matrix representing the background, and a sparse object, which may be a target or structured noise. Based on the physical model, RPCA has been utilized for principal component pursuit (PCP) for matrix low-rank recovery [4], stable PCP (SPCP) for noise-corrupted observations [5], outlier pursuit for column-sparse corruptions [6], and multi-channel RPCA [7], [8]. However, conventional RPCA often makes limited use of data priors, and the single low-rank structure cannot capture the intricate prior details of the images.

On the other hand, recent advances in deep learning have led to powerful deep denoisers for image denoising, including DnCNN [9], FFDNet [10], and DRUNet [11], to name a few. These denoisers are well-trained to capture the data priors of images and have demonstrated remarkable noise removal effects. However, they often lack physical model information and are less robust to structured noise, which can lead to degraded performance in RPCA tasks when faced with sparse impulse noise and heavily corrupted observations.

More recently, the trend of combining physical model methods with emerging deep priors has created a new paradigm for handling complex tasks in computational imaging. In this context, plug-and-play (PnP) and regularization by denoising (RED) have garnered increasing attention. For instance, PnP priors [11], [12] replace model-based proximal operators with pre-trained denoisers to recover fine textures. RED [13] constructs an explicit regularizer from the denoiser output, providing a variational interpretation that complements the implicit PnP formulation. While these approaches have been applied to various computational imaging tasks, including denoising, super-resolution, and deblurring, they have not yet been explored for more challenging RPCA tasks.

This letter focuses on leveraging the PnP framework for multi-channel RPCA, aiming to combine both physical models and deep priors. We propose an RPCA formulation that is implicitly regularized by data priors driven by deep denoisers. We apply the nonconvex ℓ2,p\ell_{2,p} norm for the structured noise removal and Schatten-pp norm for promoting the low-rank background for better separation. For the resulting problem, we introduce a PnP method embedded in a dedicated proximal distance algorithm (PDA). PnP-PDA decouples physical-model optimization and prior regularization, and uses Nesterov extrapolation for acceleration. We conducted extensive experiments on thirteen datasets, including RGB images and hyperspectral image (HSI) datasets. The results show that PnP-PDA achieves better recovery performance than conventional RPCA methods and the compared state-of-the-art deep denoisers and tensor RPCA approaches.

Background and Problem Formulation

Background of RPCA

Consider the following robust recovery signal model

Y=X+S+N,(1)\mathcal{Y} = \mathcal{X} + \mathcal{S} + \mathcal{N}, \tag*{(1)}

where Y∈RM×N×C\mathcal{Y} \in\mathbb{R}^{M \times N \times C} represents the noise corrupted multi-channel observation, X∈RM×N×C\mathcal{X} \in\mathbb{R}^{M \times N \times C} denotes the low-rank background component, S∈RM×N×C\mathcal{S} \in\mathbb{R}^{M \times N \times C} denotes the fiber-sparse outlier noise, and N∈RM×N×C\mathcal{N} \in\mathbb{R}^{M \times N \times C} denotes the additive Gaussian noise. Here, M×NM \times N denotes the data spatial size and CC denotes the number of channels.

The objective of robust multi-channel recovery is to separate and estimate X\mathcal{X} and S\mathcal{S} from Y\mathcal{Y}. This model arises in tasks such as multispectral imaging [1] and video surveillance [7], where X\mathcal{X} captures the stationary scene or spectral background and S\mathcal{S} captures small objects or anomalous pixels.

We first review the conventional RPCA formulation for the single-channel case, i.e., C=1C=1 in (1), which is given by

min⁡X,S∥X∥∗+λ∥S∥1s.t.Y=X+S,(2)\min_{\mathcal{X},\mathcal{S}} \|\mathcal{X}\|_*+\lambda\|\mathcal{S}\|_1 \quad\text{s.t.}\quad\mathcal{Y}=\mathcal{X}+\mathcal{S}, \tag*{(2)}

where X∈RM×N\mathcal{X}\in\mathbb{R}^{M\times N} denotes the low-rank component in the matrix case, and S∈RM×N\mathcal{S}\in\mathbb{R}^{M\times N} denotes the sparse outliers, λ>0\lambda>0 balances the two terms, and ∥⋅∥∗\|\cdot\|_* is the nuclear norm. This problem can be solved by accelerated proximal gradient (APG) [14], alternating direction method of multipliers (ADMM) [15], or singular value thresholding (SVT) [16].

Beyond (2), the principles of RPCA have been exploited for low-rank and sparse decomposition, including SPCP for observations with mixed Gaussian noise [5], [17], tensor decomposition methods [7], [18]. However, these conventional methods remain an issue: they often rely on physical models and typically explore less on the data priors, which may cause prior mismatch and limited recovery under severe corruptions.

Problem Formulation

To better harness both the physical model and data priors, we formulate a PnP RPCA formulation with an implicit deep prior, which reads as

min⁡X,Sγ1Φ(X)+γ2∥X∥Spp+γ3∥X∥TV+λ∥S∥2,pps.t.Y=X+S.(3)\begin{aligned} \min_{\mathcal{X},\mathcal{S}}\quad&\gamma_1\Phi(\mathcal{X})+\gamma_2\|\mathcal{X}\|_{S_p}^p+\gamma_3\|\mathcal{X}\|_{\mathrm{TV}}+\lambda\|\mathcal{S}\|_{2,p}^p\\ \text{s.t.}\quad&\mathcal{Y}=\mathcal{X}+\mathcal{S}. \tag*{(3)} \end{aligned}

In this formulation, Φ(⋅)\Phi(\cdot) is an implicit deep prior, and γ1,γ2,γ3,λ≥0\gamma_1,\gamma_2,\gamma_3,\lambda\geq0 balance the four regularizers. The ℓ2,p\ell_{2,p} norm ∥S∥2,pp=∑i,j∥S(i,j,:)∥pp\|\mathcal{S}\|_{2,p}^p=\sum_{i,j}\|\mathcal{S}(i,j,:)\|_p^p for 0<p≤10<p\leq1 enforces fiber-sparsity along the mode-3 fibers S(i,j,:)∈RC\mathcal{S}(i,j,:)\in\mathbb{R}^{C}. The tensor Schatten-pp norm is ∥X∥Spp=∑k=13∥X(k)∥Spp\|\mathcal{X}\|_{S_p}^p=\sum_{k=1}^{3}\|\mathcal{X}_{(k)}\|_{S_p}^p, where ∥X(k)∥Spp=∑mσk,mp\|\mathcal{X}_{(k)}\|_{S_p}^p=\sum_m\sigma_{k,m}^p is the matrix Schatten-pp norm with σk,m\sigma_{k,m} being the mm-th singular value of X(k)\mathcal{X}_{(k)}. Here, X(k)=unfold⁡k(X)∈Rdk×(D/dk)\mathcal{X}_{(k)}=\operatorname{unfold}_k(\mathcal{X})\in\mathbb{R}^{d_k\times(D/d_k)} denotes the tensor mode-kk matricization with dimensions (d1,d2,d3)=(M,N,C)(d_1,d_2,d_3)=(M,N,C), and D=MNCD=MNC, while the inverse operator fold⁡k\operatorname{fold}_k restores the multi-channel shape. The anisotropic TV norm is expressed in a channel-wise vectorized form as ∥X∥TV=∑c=1C(∥Dhxc∥1+∥Dvxc∥1)\|\mathcal{X}\|_{\mathrm{TV}}=\sum_{c=1}^{C}(\|D_hx_c\|_1+\|D_vx_c\|_1). Here, xc:=vec⁡(X:,:,c)∈RMNx_c:=\operatorname{vec}(\mathcal{X}_{:,:,c})\in\mathbb{R}^{MN} denotes the vectorization of the cc-th spatial channel, and Dh,Dv∈RMN×MND_h,D_v\in\mathbb{R}^{MN\times MN} are the horizontal and vertical finite-difference matrices evaluated under periodic boundary conditions.

Compared with (2), the formulation in (3) integrates the physical model with a deep prior Φ(⋅)\Phi(\cdot) through the PnP framework. Furthermore, to better promote low-rank and sparsity structures, we replace the nuclear norm and ℓ1\ell_1 norm with the nonconvex Schatten-pp and ℓ2,p\ell_{2,p} norms. However, the enhanced PnP formulation (3) presents a critical nonconvex optimization challenge, particularly due to the coupling between the deep prior Φ\Phi and the other components.

Plug-and-Play Proximal Distance Algorithm

PnP-PDA

We address the nonconvex coupling challenge with a novel PDA algorithm that reformulates the task as a consensus optimization problem. Then, the global objective can be split into tractable subproblems that allow for efficient closed-form updates. This decoupling isolates the deep prior Φ\Phi, enabling the direct embedding of a deep denoiser as a modular solver.

Our first step is to introduce the consensus constraint with X1,X2,X3=X\mathcal{X}_1,\mathcal{X}_2,\mathcal{X}_3=\mathcal{X} by

C={(X1,X2,X3,S)∣X1=X2=X3, X1+S=Y}\mathcal{C}=\{(\mathcal{X}_1,\mathcal{X}_2,\mathcal{X}_3,\mathcal{S})\mid\mathcal{X}_1=\mathcal{X}_2=\mathcal{X}_3,\ \mathcal{X}_1+\mathcal{S}=\mathcal{Y}\}.

Then, problem (3) is reformulated as

min⁡x∈Cf(x)=γ1Φ(X1)+γ2∥X2∥Spp+γ3∥X3∥TV+λ∥S∥2,pp,(4)\min_{\boldsymbol{x}\in\mathcal{C}} f(\boldsymbol{x})=\gamma_1\Phi(\mathcal{X}_1)+\gamma_2\|\mathcal{X}_2\|_{S_p}^p+\gamma_3\|\mathcal{X}_3\|_{\mathrm{TV}}+\lambda\|\mathcal{S}\|_{2,p}^p, \tag*{(4)}

where x=(X1,X2,X3,S)\boldsymbol{x}=(\mathcal{X}_1,\mathcal{X}_2,\mathcal{X}_3,\mathcal{S}). Next, we transform the consensus constraint with a soft constraint, which is posed as

Fρ(x)=f(x)+ρ2dist⁡2(x,C),(5)F_\rho(\boldsymbol{x})=f(\boldsymbol{x})+\frac{\rho}{2}\operatorname{dist}^2(\boldsymbol{x},\mathcal{C}), \tag*{(5)}

where dist⁡(x,C)=min⁡y∈C∥x−y∥F\operatorname{dist}(\boldsymbol{x},\mathcal{C})=\min_{\boldsymbol{y}\in\mathcal{C}}\|\boldsymbol{x}-\boldsymbol{y}\|_F is the distance function and ρ>0\rho>0 is a penalty parameter. We majorize dist⁡2(x,C)\operatorname{dist}^2(\boldsymbol{x},\mathcal{C}) at xt\boldsymbol{x}^t using a quadratic surrogate

dist⁡2(x,C)≤∥x−PC(xt)∥F2,(6)\operatorname{dist}^2(\boldsymbol{x},\mathcal{C})\leq\|\boldsymbol{x}-P_{\mathcal{C}}(\boldsymbol{x}^t)\|_F^2, \tag*{(6)}

with equality at x=xt\boldsymbol{x}=\boldsymbol{x}^t, where PCP_{\mathcal{C}} denotes the projection onto C\mathcal{C} that takes the form PC(xt)=(PCX,PCX,PCX,PCS)P_{\mathcal{C}}(\boldsymbol{x}^t)=(P_{\mathcal{C}}^{\mathcal{X}},P_{\mathcal{C}}^{\mathcal{X}},P_{\mathcal{C}}^{\mathcal{X}},P_{\mathcal{C}}^{\mathcal{S}}) with

PCX=14(∑i=13Xit+Y−St),PCS=Y−PCX.(7)P_{\mathcal{C}}^{\mathcal{X}}=\frac{1}{4}\left(\sum_{i=1}^{3}\mathcal{X}_i^t+\mathcal{Y}-\mathcal{S}^t\right),\qquad P_{\mathcal{C}}^{\mathcal{S}}=\mathcal{Y}-P_{\mathcal{C}}^{\mathcal{X}}. \tag*{(7)}

Under this majorization, we propose the following PDA algorithm where at iteration tt it takes the form

xt+1=arg⁡min⁡xf(x)+ρ2∑i=13∥Xi−PCX∥F2+ρ2∥S−PCS∥F2.(8)\boldsymbol{x}^{t+1}=\arg\min_{\boldsymbol{x}}f(\boldsymbol{x})+\frac{\rho}{2}\sum_{i=1}^{3}\|\mathcal{X}_i-P_{\mathcal{C}}^{\mathcal{X}}\|_F^2+\frac{\rho}{2}\|\mathcal{S}-P_{\mathcal{C}}^{\mathcal{S}}\|_F^2. \tag*{(8)}

Since f(x)f(\boldsymbol{x}) is separable, problem (8) splits into four independent subproblems for X1,X2,X3\mathcal{X}_1,\mathcal{X}_2,\mathcal{X}_3, and S\mathcal{S}, each admitting a form of proximal operator and can be efficiently solved.

  1. Proximal Mapping for X1\mathcal{X}_1: The X1\mathcal{X}_1 subproblem reads

min⁡X1 γ1Φ(X1)+ρ2∥X1−PCX∥F2.(9)\min_{\mathcal{X}_1}\ \gamma_1\Phi(\mathcal{X}_1)+\frac{\rho}{2}\|\mathcal{X}_1-P_{\mathcal{C}}^{\mathcal{X}}\|_F^2. \tag*{(9)}

which amounts to derive the proximal mapping prox⁡(γ1/ρ)Φ(PCX)\operatorname{prox}_{(\gamma_1/\rho)\Phi}(P_{\mathcal{C}}^{\mathcal{X}}). Since Φ\Phi is an implicit data prior, its proximal subproblem can be addressed by a denoiser under the PnP framework [19]:

X1t+1=Denoiser⁡(PCX,σt),(10)\mathcal{X}_1^{t+1}=\operatorname{Denoiser}(P_{\mathcal{C}}^{\mathcal{X}},\sigma_t), \tag*{(10)}

where σt=γ1/ρ\sigma_t=\sqrt{\gamma_1/\rho} is the noise level parameter.

  1. Proximal Mapping for X2\mathcal{X}_2: The X2\mathcal{X}_2 subproblem is

min⁡X2 γ2∥X2∥Spp+ρ2∥X2−PCX∥F2.(11)\min_{\mathcal{X}_2}\ \gamma_2\|\mathcal{X}_2\|_{S_p}^p+\frac{\rho}{2}\|\mathcal{X}_2-P_{\mathcal{C}}^{\mathcal{X}}\|_F^2. \tag*{(11)}

Since the X2\mathcal{X}_2 subproblem sums over three mode unfoldings coupled through the same tensor variable, we approximate it using a mode-wise decoupling-and-averaging strategy with three matrix subproblems for k∈{1,2,3}k\in\{1,2,3\}, written as

min⁡X2(k) γ2∥X2(k)∥Spp+ρ2∥X2(k)−PC(k)X∥F2.(12)\min_{\mathcal{X}_{2(k)}}\ \gamma_2\|\mathcal{X}_{2(k)}\|_{S_p}^p+\frac{\rho}{2}\|\mathcal{X}_{2(k)}-P_{\mathcal{C}(k)}^{\mathcal{X}}\|_F^2. \tag*{(12)}

Let PC(k)X=UkΣkVkTP_{\mathcal{C}(k)}^{\mathcal{X}}=U_k\Sigma_kV_k^{\mathsf{T}} be the SVD of the mode-kk matricization of PCXP_{\mathcal{C}}^{\mathcal{X}}, where σk,m\sigma_{k,m} denotes its mm-th singular value. By the unitary invariance of the Schatten-pp and Frobenius norms, the optimal X2(k)\mathcal{X}_{2(k)} shares the singular vectors UkU_k and VkV_k of PC(k)X\mathcal{P}^{\mathcal{X}}_{C(k)} [16], and the optimization reduces to dkd_k independent scalar problems for the sought singular values σ~m≥0\tilde{\sigma}_m \ge0

min⁡σ~m≥0γ2σ~mp+ρ2(σ~m−σk,m)2,m=1,…,dk.(13)\min_{\tilde{\sigma}_m \ge0} \gamma_2 \tilde{\sigma}_m^p + \frac{\rho}{2}(\tilde{\sigma}_m-\sigma_{k,m})^2,\quad m=1,\ldots,d_k. \tag*{(13)}

Since 0<p<10 < p < 1, the term σ~mp=(σ~m2)p/2\tilde{\sigma}_m^p = (\tilde{\sigma}_m^2)^{p/2} is non-convex. Following the reweighted least squares (RLS) strategy [20], we majorize this term at the current iterate by the first-order upper bound σ~mp≤wk(m)σ~m2+ck(m)\tilde{\sigma}_m^p \le w_k^{(m)}\tilde{\sigma}_m^2+c_k^{(m)}, where ck(m)c_k^{(m)} is a constant and the scalar weight is

wk(m)=p2((σmt)2+ϵ)p−22,(14)w_k^{(m)}=\frac{p}{2}\left((\sigma_m^t)^2+\epsilon\right)^{\frac{p-2}{2}}, \tag*{(14)}

with σmt\sigma_m^t denoting the corresponding singular value of the previous iterate X2(k)t\mathcal{X}_{2(k)}^t and ϵ>0\epsilon> 0 ensuring numerical stability. Substituting this majorizer back into (13) and setting the derivative to zero yields the closed-form reweighted shrinkage

σ~m=ρρ+2γ2wk(m)σk,m.(15)\tilde{\sigma}_m=\frac{\rho}{\rho+2\gamma_2w_k^{(m)}}\sigma_{k,m}. \tag*{(15)}

Averaging the three mode estimates and refolding yields the multi-channel update

X2t+1=13∑k=13fold⁡k(UkΣ~kVkT),(16)\mathcal{X}_2^{t+1}=\frac{1}{3}\sum_{k=1}^{3}\operatorname{fold}_k\left(U_k\widetilde{\Sigma}_kV_k^{\mathsf{T}}\right), \tag*{(16)}

where Σ~k=diag⁡(σ~1,…,σ~dk)\widetilde{\Sigma}_k=\operatorname{diag}(\tilde{\sigma}_1,\ldots,\tilde{\sigma}_{d_k}). Within each PDA iteration, the RLS performs a single reweighting step.

Proximal Mapping for X3\mathcal{X}_3

The X3\mathcal{X}_3-subproblem is

min⁡X3∑c=1C[γ3∥X3,c∥TV+ρ2∥X3,c−PC,cX∥F2].(17)\min_{\mathcal{X}_3}\sum_{c=1}^{C}\left[\gamma_3\|\mathcal{X}_{3,c}\|_{\mathrm{TV}}+\frac{\rho}{2}\|\mathcal{X}_{3,c}-\mathcal{P}^{\mathcal{X}}_{C,c}\|_F^2\right]. \tag*{(17)}

where X3,c,PC,cX∈RM×N\mathcal{X}_{3,c},\mathcal{P}^{\mathcal{X}}_{C,c}\in\mathbb{R}^{M\times N} denote the cc-th channel of X3\mathcal{X}_3 and PX\mathcal{P}^{\mathcal{X}}, respectively. Since the problem is independent across channels, we describe the procedure for each channel. Since DhD_h and DvD_v couple the entries of xcx_c, we adopt Fast Total Variation deconvolution (FTVd) [21] by introducing auxiliary variables gh=Dhxcg_h=D_hx_c and gv=Dvxcg_v=D_vx_c and relaxing the equality constraints with a quadratic penalty of weight β\beta,

min⁡xc,gh,gvγ3(∥gh∥1+∥gv∥1)+ρ2∥xc−vec⁡(PC,cX)∥2+β2∥gh−Dhxc∥2+β2∥gv−Dvxc∥2.(18)\min_{x_c,g_h,g_v}\gamma_3(\|g_h\|_1+\|g_v\|_1)+\frac{\rho}{2}\|x_c-\operatorname{vec}(\mathcal{P}^{\mathcal{X}}_{C,c})\|^2+\frac{\beta}{2}\|g_h-D_hx_c\|^2+\frac{\beta}{2}\|g_v-D_vx_c\|^2. \tag*{(18)}

We choose β=ρ\beta=\rho. For any given xcx_c, the solutions for ghg_h and gvg_v is given by soft-thresholding

g^h=Tγ3/ρ(Dhxc),g^v=Tγ3/ρ(Dvxc),(19)\hat{g}_h=\mathcal{T}_{\gamma_3/\rho}(D_hx_c),\qquad\hat{g}_v=\mathcal{T}_{\gamma_3/\rho}(D_vx_c), \tag*{(19)}

where Tτ(x)=sgn⁡(x)max⁡(∣x∣−τ,0)\mathcal{T}_{\tau}(x)=\operatorname{sgn}(x)\max(|x|-\tau,0). Substituting g^h,g^v\hat{g}_h,\hat{g}_v back into (18) and setting the gradient with respect to xcx_c to zero gives the normal equation Axc=bAx_c=b, where

A=I+DhTDh+DvTDv,b=vec⁡(PC,cX)+DhTg^h+DvTg^v.A=I+D_h^{\mathsf{T}}D_h+D_v^{\mathsf{T}}D_v,\quad b=\operatorname{vec}(\mathcal{P}^{\mathcal{X}}_{C,c})+D_h^{\mathsf{T}}\hat{g}_h+D_v^{\mathsf{T}}\hat{g}_v.

Under periodic boundary conditions, DhD_h and DvD_v are circulant and diagonalized by the DFT matrix FF as Dh=FHΛhFD_h=F^{H}\Lambda_hF and Dv=FHΛvFD_v=F^{H}\Lambda_vF. Then, the closed-form solution is

xc=FHΛ−1Fb,(20)x_c=F^{H}\Lambda^{-1}Fb, \tag*{(20)}

Algorithm 1 PnP-PDA for RPCA

Require: Observation Y\mathcal{Y}, parameters γ1,γ2,γ3,λ,p\gamma_1,\gamma_2,\gamma_3,\lambda,p, smoothing factor ϵ\epsilon, scaling factor κ>1\kappa>1, and ρmax⁡\rho_{\max}.

  1. Initialize: X10=MedianFilter⁡(Y)\mathcal{X}_1^0=\operatorname{MedianFilter}(\mathcal{Y}); S0=0\mathcal{S}^0=0; ρ0>0\rho^0>0.

  2. repeat

  3. Compute extrapolated points Z1t,ZSt\mathcal{Z}_1^t,\mathcal{Z}_S^t via (23);

  4. Compute consensus projection PX,PS\mathcal{P}^{\mathcal{X}},\mathcal{P}^{\mathcal{S}} via (7);

  5. Update X1t+1\mathcal{X}_1^{t+1} via deep denoiser (10);

  6. Update X2t+1\mathcal{X}_2^{t+1} via RLS singular value shrinkage (16);

  7. Update X3t+1\mathcal{X}_3^{t+1} via TV proximal operator (20);

  8. Update St+1\mathcal{S}^{t+1} via RLS tubal shrinkage (22);

  9. Update penalty: ρt+1←min⁡(κρt,ρmax⁡)\rho^{t+1}\leftarrow\min(\kappa\rho^t,\rho_{\max});

  10. until convergence

  11. Return PCX,PCS=Y−PCX\mathcal{P}^{\mathcal{X}}_C,\mathcal{P}^{\mathcal{S}}_C=\mathcal{Y}-\mathcal{P}^{\mathcal{X}}_C.

where Λ=I+ΛhHΛh+ΛvHΛv\Lambda=I+\Lambda_h^H\Lambda_h+\Lambda_v^H\Lambda_v is a constant diagonal matrix independent of ρ\rho and tt, precomputed once and reused across all CC channels.

Proximal Mapping for S\mathcal{S}

The S\mathcal{S} subproblem reads as

min⁡Sλ∥S∥p,2p+ρ2∥S−PCS∥F2.(21)\min_{\mathcal{S}}\lambda\|\mathcal{S}\|_{p,2}^p+\frac{\rho}{2}\|\mathcal{S}-\mathcal{P}^{\mathcal{S}}_C\|_F^2. \tag*{(21)}

Applying the same RLS scheme as in the X2\mathcal{X}_2 subproblem yields the closed-form reweighted tubal shrinkage

St+1(i,j,:)=ρρ+2λωi,jPCS(i,j,:),(22)\mathcal{S}^{t+1}(i,j,:)=\frac{\rho}{\rho+2\lambda\omega_{i,j}}\mathcal{P}^{\mathcal{S}}_C(i,j,:), \tag*{(22)}

where ωi,j=p2(∥St(i,j,:)∥2+ϵ)(p−2)/2\omega_{i,j}=\frac{p}{2}\left(\|\mathcal{S}^t(i,j,:)\|^2+\epsilon\right)^{(p-2)/2} is the fiber-wise weight at spatial location (i,j)(i,j).

PDA Acceleration and Complexity Analysis

We introduce Nesterov-type extrapolation to accelerate numerical convergence, which has been shown to effectively accelerate convergence in many optimization algorithms [22, 23, 24]. At iteration tt, the extrapolated point is calculated by

Zit=Xit+t−1t+2(Xit−Xit−1),i∈{1,2,3},(23)\mathcal{Z}_i^t=\mathcal{X}_i^t+\frac{t-1}{t+2}(\mathcal{X}_i^t-\mathcal{X}_i^{t-1}),\quad i\in\{1,2,3\}, \tag*{(23)}

and similarly ZSt\mathcal{Z}_S^t for St\mathcal{S}^t. The PDA update is then performed at the extrapolated point by replacing Xit\mathcal{X}_i^t with Zit\mathcal{Z}_i^t and St\mathcal{S}^t with ZSt\mathcal{Z}_S^t in the consensus projection (7), while the four proximal subproblems (10)–(22) remain unchanged. We also adopt a homotopy strategy for ρ\rho by initializing it with a small value and increasing it via ρt+1=min⁡(κρt,ρmax⁡)\rho^{t+1}=\min(\kappa\rho^t,\rho_{\max}) with κ>1\kappa>1, which gradually enforces the consensus constraints. The complete procedure is summarized in Algorithm 1. The algorithm complexity analysis is provided in the supplementary material.

Simulation Results

Simulation Settings and Benchmarks

We evaluate PnP-PDA on thirteen images, including ten RGB images and three HSI datasets: PaviaU (103 bands), Samson (156 bands), and Jasper (198 bands). All data are tested under two corruption ratios (10% and 20%) and Gaussian noise (σ∈{0.10,0.20}\sigma\in\{0.10,0.20\}). Corrupted fibers are perturbed by i.i.d. uniform noise in [−0.8,0.8][-0.8,0.8]. The PnP framework can plug different pre-trained denoisers for different datasets, including DRUNet [11] for RGB images and DPHSIR [27] for hyperspectral images. We use PSNR, SSIM [28], and LPIPS [29] metrics for RGB images, and band-averaged MPSNR, MSSIM [30] metrics for HSI. The definitions of these metrics are provided in the Information for Reproducibility.

We compare PnP-PDA against seven benchmarks and one ablation baseline. Two are conventional RPCA methods: Inexact ALM (IALM) [15] and SPCP [5]. Three are state-of-the-art deep denoisers: DnCNN [9], FFDNet [10], and DRUNet [11]. Two are state-of-the-art tensor RPCA methods: Unfolded ScaledGD (U-SGD) [25] and t-CTV [26]. We added an ablation study as a baseline by examining the deep learning prior in our PnP-PDA framework.

Performance Comparison

Figure 1 compares results on Girl and Bell at σ=0.10\sigma= 0.10. Conventional RPCA methods, the ablation variant, and state-of-the-art tensor RPCA methods produce over-smoothed results, and deep denoisers leave structured artifacts, whereas PnP-PDA removes corruptions while preserving textures and edges. Extensive RGB and HSI visual results under a wider range of noise levels are provided in the supplementary material.

Visual results on “Girl” and “Bell”

Figure 1. Visual results on “Girl” and “Bell”: Original, Corruption, IALM, SPCP, DnCNN, FFDNet, DRUNet, U-SGD, t-CTV, Ablation, and PnP-PDA.

Table I reports the results on ten diverse RGB images at σ=0.10\sigma= 0.10. PnP-PDA achieves better performance than the state-of-the-art tensor RPCA baselines across all tests. The improvement over the ablation (without PnP) shows the contribution of the deep prior in our framework. Overall, PnP-PDA demonstrates consistently competitive quantitative performance on all ten RGB images.

Method (PSNR/SSIM/LPIPS)sp\mathrm{sp}PeppersGirlBellHousePlaneChildkodim23kodim24kodim04kodim15
IALM [15]0.113.27/0.3178/0.954714.94/0.3883/0.801316.95/0.2421/1.012314.66/0.3353/0.911316.17/0.3885/0.797217.41/0.3971/0.883414.11/0.3169/1.124817.20/0.5077/0.665916.14/0.3486/0.996315.35/0.3772/0.8045
0.212.94/0.2770/1.024714.04/0.3221/0.905715.43/0.1898/1.101913.71/0.2784/1.014814.60/0.3243/0.933216.74/0.3401/0.961413.48/0.2616/1.169915.91/0.4171/0.819815.24/0.2869/1.101614.27/0.3227/0.8925
SPCP [5]0.112.49/0.2929/1.014714.67/0.3766/0.818216.66/0.2342/1.026214.38/0.3296/0.914016.04/0.3846/0.806116.67/0.3767/0.899313.61/0.2952/1.133416.49/0.4908/0.687115.95/0.3447/1.010015.02/0.3712/0.8155
0.212.22/0.2547/1.087213.82/0.3129/0.920615.22/0.1857/1.109513.50/0.2751/1.020914.47/0.3210/0.940816.05/0.3245/0.981013.06/0.2457/1.177115.37/0.4059/0.832915.05/0.2839/1.108713.97/0.3175/0.9002
DnCNN [9]0.125.83/0.7587/0.136927.26/0.8013/0.120629.94/0.8274/0.170928.19/0.8090/0.065126.21/0.7953/0.124225.65/0.6860/0.244027.02/0.7730/0.114125.70/0.7694/0.134127.29/0.7603/0.140225.99/0.7419/0.1792
0.224.26/0.7060/0.179525.83/0.7539/0.149028.48/0.7817/0.211326.56/0.7633/0.105224.60/0.7320/0.173523.75/0.6002/0.336425.51/0.7182/0.186024.49/0.7100/0.189625.71/0.7059/0.209323.98/0.6706/0.2347
FFDNet [10]0.123.42/0.6101/0.331323.66/0.6110/0.340423.93/0.4312/0.566123.81/0.5495/0.310323.63/0.6252/0.276523.38/0.5650/0.476123.66/0.6020/0.351822.90/0.6522/0.312323.37/0.5614/0.457423.61/0.6043/0.3600
0.219.93/0.4429/0.589219.78/0.4293/0.541319.70/0.2493/0.883719.76/0.3692/0.589119.73/0.4347/0.526719.75/0.3754/0.808219.88/0.4289/0.575919.59/0.4813/0.603619.65/0.3747/0.759219.87/0.4157/0.5901
DRUNet [11]0.121.48/0.5332/0.369421.45/0.5049/0.371621.52/0.3157/0.609121.61/0.4552/0.337521.58/0.5366/0.316721.30/0.4719/0.538221.36/0.5012/0.403120.87/0.5453/0.339721.27/0.4580/0.491621.63/0.5217/0.4266
0.218.10/0.3684/0.596617.77/0.3425/0.552317.66/0.1809/0.878017.84/0.2994/0.578718.06/0.3655/0.536017.96/0.2969/0.816717.86/0.3418/0.581417.50/0.3735/0.633117.68/0.2852/0.755918.18/0.3461/0.6105
U-SGD [25]0.122.50/0.5568/0.368224.51/0.6560/0.422926.56/0.5751/0.518223.37/0.4939/0.372322.04/0.5413/0.349124.05/0.6072/0.505922.65/0.5482/0.391020.88/0.5784/0.380122.82/0.5406/0.551522.43/0.551?
0.221.??22.??25.??22.??20.??23.??21.??20.??21.??21.??
t-CTV [26]0.124.36/0.6355/0.248826.00/0.6766/0.226926.67/0.5196/0.435225.32/0.5662/0.225123.46/0.5930/0.288425.68/0.6352/0.322024.33/0.6355/0.269222.58/0.6779/0.296424.45/0.6068/0.314824.09/0.6124/0.2699
0.223.69/0.6019/0.276525.01/0.6339/0.284725.78/0.4750/0.494024.50/0.5540/0.267522.81/0.5559/0.317824.89/0.5974/0.381923.66/0.6022/0.312421.97/0.6403/0.339723.76/0.5691/0.364523.54/0.5831/0.2941
Without PnP (ablation)0.124.21/0.6155/0.252726.17/0.6959/0.285527.48/0.6481/0.463025.02/0.5659/0.253623.46/0.5871/0.291925.34/0.6232/0.334524.20/0.6324/0.304422.54/0.6601/0.269824.46/0.6104/0.336323.89/0.5906/0.2759
0.223.48/0.5976/0.333525.61/0.6827/0.326127.12/0.6532/0.494224.44/0.5576/0.279822.84/0.5710/0.347024.68/0.6032/0.418023.64/0.6183/0.350421.96/0.6304/0.312123.96/0.5941/0.405223.11/0.5755/0.3410
PnP-PDA0.128.09/0.8358/0.049230.36/0.8654/0.057332.75/0.9171/0.036431.02/0.8485/0.027527.01/0.8442/0.050229.00/0.8328/0.059028.29/0.8382/0.051324.96/0.7900/0.109627.99/0.7869/0.133028.30/0.8271/0.0996
0.227.01/0.8190/0.054329.55/0.8558/0.080031.47/0.9008/0.046129.76/0.8336/0.030725.46/0.8248/0.073227.84/0.8126/0.089726.80/0.8155/0.071223.80/0.7264/0.106826.91/0.7605/0.129926.95/0.8099/0.0975

Table I. Comparison of recovery methods on ten RGB benchmark images at σ=0.10\sigma= 0.10 under 10% and 20% corruption.

Table II reports the results on three HSI datasets with 103, 156, and 198 spectral bands. PnP-PDA achieves the highest MPSNR and MSSIM in all 12 test conditions. It consistently outperforms the conventional RPCA methods, state-of-the-art deep denoisers, and other RPCA methods, and the ablation variant. These results support the effectiveness of PnP-PDA for HSI recovery.

Method (MPSNR/MSSIM)sp=0.1\mathrm{sp}=0.1, PaviaU 0.10PaviaU 0.20Samson 0.10Samson 0.20Jasper 0.10Jasper 0.20sp=0.2\mathrm{sp}=0.2, PaviaU 0.10PaviaU 0.20Samson 0.10Samson 0.20Jasper 0.10Jasper 0.20
IALM [15]27.86/0.76925.22/0.67030.98/0.79528.34/0.70729.62/0.82327.05/0.73726.18/0.69224.33/0.62129.30/0.73527.44/0.67027.87/0.76526.06/0.696
SPCP [5]27.85/0.76825.17/0.66730.92/0.79428.25/0.70429.60/0.82227.03/0.73726.15/0.69124.27/0.61829.24/0.73327.35/0.66727.84/0.76526.04/0.696
DnCNN [9]25.80/0.64724.25/0.61925.48/0.56823.48/0.58025.37/0.61423.52/0.58023.73/0.54323.81/0.55123.70/0.45522.02/0.49923.38/0.52222.21/0.518
FFDNet [10]27.21/0.70124.81/0.65828.09/0.75223.94/0.65926.93/0.67423.97/0.62726.02/0.69523.16/0.57626.05/0.71422.43/0.58025.63/0.66122.48/0.559
DRUNet [11]27.58/0.73024.44/0.62928.28/0.76023.82/0.62827.18/0.69223.64/0.59525.21/0.65025.14/0.57925.06/0.68122.61/0.61125.04/0.62522.58/0.536
U-SGD [25]30.35/0.79526.85/0.62729.18/0.62125.86/0.46030.20/0.77426.95/0.63829.12/0.74726.12/0.58927.14/0.53724.72/0.41528.11/0.70425.73/0.589
t-CTV [26]27.42/0.66522.80/0.42428.80/0.62124.18/0.39628.19/0.70223.37/0.47426.58/0.62523.23/0.39528.05/0.58523.62/0.37027.33/0.66522.76/0.444
Ablation29.69/0.82222.15/0.39529.80/0.78926.26/0.55230.80/0.84023.85/0.50028.89/0.78921.06/0.34429.28/0.77824.31/0.44029.85/0.81222.40/0.427
PnP-PDA31.34/0.86429.38/0.81233.86/0.87629.77/0.82031.33/0.85228.08/0.75730.46/0.83028.80/0.79332.24/0.83628.83/0.79530.09/0.80427.54/0.752

Table II. Comparison of HSI recovery methods under 10% and 20% corruption at Gaussian noise levels σ=0.10\sigma= 0.10 and σ=0.20\sigma= 0.20.

Conclusion

This letter presents a PnP method for RPCA that integrates both the physical model and deep priors. Our proposed PnP-PDA algorithm transforms the problem into subproblems for evaluating proximal operators, utilizing closed-form updates and deep denoiser processing. Extensive experiments on diverse RGB and HSI datasets demonstrate the efficacy of our PnP-PDA method.

References

  1. [1]M. M. Salut and D. V. Anderson, “Tensor robust CUR for compression and denoising of hyperspectral data,” IEEE Access, vol. 11, pp. 77 492–77 505, 2023.DOI
  2. [2]J. Wright, A. Y. Yang, A. Ganesh, S. S. Sastry, and Y. Ma, “Robust face recognition via sparse representation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 31, no. 2, pp. 210–227, 2009.DOI
  3. [3]Z. Han, S. Zhang, Z. Liu, Y. Wang, J. Yao, and Y. Wang, “Tensor robust principal component analysis with side information: Models and applications,” IEEE Trans. Circuits Syst. Video Technol., vol. 33, no. 8, pp. 3713–3725, 2023.DOI
  4. [4]E. J. Candès, X. Li, Y. Ma, and J. Wright, “Robust principal component analysis?” J. ACM, vol. 58, no. 3, pp. 1–37, 2011.arxiv.org/abs/0912.3599
  5. [5]Z. Zhou, X. Li, J. Wright, E. J. Candès, and Y. Ma, “Stable principal component pursuit,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), 2010, pp. 1518–1522.arxiv.org/abs/1001.2363
  6. [6]H. Xu, C. Caramanis, and S. Sanghavi, “Robust PCA via outlier pursuit,” IEEE Trans. Inf. Theory, vol. 58, no. 5, pp. 3047–3064, 2012.arxiv.org/abs/1010.4237
  7. [7]C. Lu, J. Feng, Y. Chen, W. Liu, Z. Lin, and S. Yan, “Tensor robust principal component analysis with a new tensor nuclear norm,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 42, no. 4, pp. 925–938, 2019.DOI
  8. [8]B. Huang, C. Mu, D. Goldfarb, and J. Wright, “Provable low-rank tensor recovery,” Optimization Online, Tech. Rep. 4252, 2014.
  9. [9]K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,” IEEE Trans. Image Process., vol. 26, no. 7, pp. 3142–3155, 2017.DOI
  10. [10]K. Zhang, W. Zuo, and L. Zhang, “FFDNet: Toward a fast and flexible solution for CNN-based image denoising,” IEEE Trans. Image Process., vol. 27, no. 9, pp. 4608–4622, 2018.DOI
  11. [11]K. Zhang, Y. Li, W. Zuo, L. Zhang, L. V. Gool, and R. Timofte, “Plug-and-play image restoration with deep denoiser prior,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 10, pp. 6360–6376, 2022.arxiv.org/abs/2008.13751
  12. [12]S. V. Venkatakrishnan, C. A. Bouman, and B. Wohlberg, “Plug-and-play priors for model based reconstruction,” in Proc. IEEE Global Conf. Signal Inf. Process. (GlobalSIP), 2013, pp. 945–948.
  13. [13]Y. Romano, M. Elad, and P. Milanfar, “The little engine that could: Regularization by denoising (RED),” SIAM J. Imaging Sci., vol. 10, no. 4, pp. 1804–1844, 2017.arxiv.org/abs/1611.02862
  14. [14]A. Ganesh, Z. Lin, J. Wright, L. Wu, M. Chen, and Y. Ma, “Fast algorithms for recovering a corrupted low-rank matrix,” in Proc. 3rd IEEE Int. Workshop Comput. Adv. Multi-Sensor Adapt. Process. (CAMSAP), 2009, pp. 213–216.DOI
  15. [15]Z. Lin, M. Chen, and Y. Ma, “The augmented Lagrange multiplier method for exact recovery of corrupted low-rank matrices,” arXiv preprint arXiv:1009.5055, 2010.DOI
  16. [16]J.-F. Cai, E. J. Candès, and Z. Shen, “A singular value thresholding algorithm for matrix completion,” SIAM J. Optim., vol. 20, no. 4, pp. 1956–1982, 2010.arxiv.org/abs/0810.3286
  17. [17]T. Bouwmans and E. H. Zahzah, “Robust PCA via principal component pursuit: A review for a comparative evaluation in video surveillance,” Comput. Vis. Image Underst., vol. 122, pp. 22–34, 2014.
  18. [18]A. Cichocki, D. Mandic, L. De Lathauwer, G. Zhou, Q. Zhao, C. Caiafa, and H. A. Phan, “Tensor decompositions for signal processing applications: From two-way to multiway component analysis,” IEEE Signal Process. Mag., vol. 32, no. 2, pp. 145–163, 2015.DOI
  19. [19]Y. Zhu, K. Zhang, J. Liang, J. Cao, B. Wen, R. Timofte, and L. V. Gool, “Denoising diffusion models for plug-and-play image restoration,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops, 2023, pp. 1219–1229.arxiv.org/abs/2305.08995
  20. [20]X. Fu, K. Huang, W.-K. Ma, N. D. Sidiropoulos, and R. Bro, “Joint tensor factorization and outlying slab suppression with applications,” IEEE Trans. Signal Process., vol. 63, no. 23, pp. 6315–6328, 2015.DOI
  21. [21]Y. Wang, J. Yang, W. Yin, and Y. Zhang, “A new alternating minimization algorithm for total variation image reconstruction,” SIAM J. Imaging Sci., vol. 1, no. 3, pp. 248–272, 2008.DOI
  22. [22]K. L. Keys, H. Zhou, and K. Lange, “Proximal distance algorithms: Theory and practice,” Journal of Machine Learning Research, vol. 20, no. 66, pp. 1–38, 2019.pubmed.ncbi.nlm.nih.gov/31649491
  23. [23]Y. Nesterov, Introductory Lectures on Convex Optimization: A Basic Course. Springer, 2013.
  24. [24]M. Shao, W.-K. Ma, J. Liu, and Z. Huang, “Accelerated and deep expectation maximization for one-bit MIMO-OFDM detection,” IEEE Trans. Signal Process., vol. 72, pp. 1094–1113, 2024.arxiv.org/abs/2210.03888
  25. [25]H. Dong, M. Shah, S. Donegan, and Y. Chi, “Deep unfolded tensor robust PCA with self-supervised learning,” in Proc. IEEE Int. Conf. Acoust., Speech Signal Process., 2023, pp. 1–5.arxiv.org/abs/2212.11346
  26. [26]H. Wang, J. Peng, W. Qin, J. Wang, and D. Meng, “Guaranteed tensor recovery fused low-rankness and smoothness,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 9, pp. 10990–11007, 2023.DOI
  27. [27]Z. Lai, K. Wei, and Y. Fu, “Deep plug-and-play prior for hyperspectral image restoration,” Neurocomputing, vol. 481, pp. 281–293, 2022.DOI
  28. [28]Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,” IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612, 2004.pubmed.ncbi.nlm.nih.gov/15376593
  29. [29]R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 586–595.arxiv.org/abs/1801.03924
  30. [30]C. Peng, Y. Chen, L. Kang, H. Chen, and Q. Cheng, “Hyperspectral image denoising using nonconvex local low-rank and sparse separation with spatial-spectral total variation regularization,” IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–17, 2022.DOI

Paper details

vlnc:W_s4mqg4wc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC0-1.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents