PaperlayerView exact revision on arXiv

Source-linked AI summary

Tensor Orthogonal Subspace Split: Theory and Applications

Jifei Miao, Juan Han, Michael K. Ng, Kit Ian Kou

arXiv:2608.29638v1math.NA

TL;DR

The paper addresses the limited explicit modeling of subspace-consistent structure and orthogonal deviations in high-order tensors. It proposes TOSS and rank-one TOSS, establishes their properties, and applies rank-one TOSS to two restoration and modeling tasks. Experiments validate the methods, with the color-video approach achieving the best performance across all reported metrics and videos.

  • Problem

    Existing tensor decomposition and low-rank methods do not explicitly capture a prescribed-mode dominant subspace component together with its orthogonal residual counterpart.

  • Method

    TOSS extends orthogonal projection to tensors through fiber-wise mode projections, while rank-one TOSS imposes a separable structure along one dominant direction and supports ADMM-based application models.

  • Results

    The color-video method achieves the best performance across all compared methods and metrics on all six videos, with CQM exceeding the second-best method by an average of approximately 3.1%.

  • Takeaways & Limitations

    TOSS provides a principled tensor mechanism for separating dominant subspace-aligned information from orthogonal residual variations, with demonstrated practical value in two representative applications.

Abstract

from arXiv · show

Tensor representations have emerged as a fundamental paradigm for modeling multidimensional data by preserving intrinsic correlations across multiple modes. This paper proposes a novel theoretical framework, termed Tensor Orthogonal Subspace Split (TOSS), which explicitly splits a tensor, along a prescribed mode, into two orthogonal components: a dominant component lying in a prescribed subspace and a residual component lying in the corresponding orthogonal complement. We first present the general formulation of TOSS and systematically investigate its fundamental properties. As an important and practically meaningful special case, we further introduce the rank-one TOSS, which imposes a separable rank-one structure along the splitting mode and admits a clear geometric interpretation. This formulation naturally captures dominant consistent patterns while effectively isolating orthogonal residual component. The proposed framework establishes a unified theoretical foundation for tensor-domain orthogonal split and opens new avenues for structured tensor modeling across diverse applications. Building upon the developed TOSS theory, hyperspectral image restoration and color video background modeling are considered as two representative tasks, for which corresponding optimization models are formulated. Efficient algorithms are developed to solve the resulting problems. Extensive experimental results validate the effectiveness and superiority of the proposed approaches.

I. INTRODUCTION

The paper introduces TOSS to extend orthogonal projection from vectors to tensors, explicitly separating prescribed-subspace structure from orthogonal residuals. It develops rank-one TOSS and applies it to hyperspectral image restoration and color video background modeling.

  • Motivation: Existing tensor methods preserve multidimensional correlations and exploit low-rank structure, but do not explicitly capture mode-wise dominant consistency and orthogonal deviation.The paper motivates a framework beyond compact representation or low-rank approximation.
  • Tensor Orthogonal Subspace Split: TOSS extends classical orthogonal projection to tensors by inducing fiber-wise projections along a prescribed mode.The split is exact and energy-preserving, producing a dominant subspace component and an orthogonal residual component.
  • Rank-one TOSS: Rank-one TOSS restricts the underlying subspace to one dominant direction, yielding a separable rank-one structure that captures consistent patterns and isolates orthogonal deviations.Its geometric interpretation makes it suitable for data with strong mode-wise consistency.
  • Applications: Rank-one TOSS supports optimization models for hyperspectral image restoration and color video background modeling, solved using efficient scaled ADMM algorithms.The paper reports extensive experiments demonstrating the effectiveness and superiority of the resulting methods.
  • Modeling Perspective: The framework separately models dominant and residual components while jointly exploiting them for high-dimensional tensor processing tasks.This modeling mechanism is presented as a response to the limited explicit separation available in existing methods.

II. NOTATION AND PRELIMINARIES

This section establishes notation for tensors, fibers, mode-n products, inner products, norms, and total-variation regularizers used throughout the paper.

  • Notation: Scalars, vectors, matrices, and tensors use lowercase, bold lowercase, bold uppercase, and calligraphic symbols, respectively.The identity matrices are denoted by I and I_n, while (·)^⊤ denotes matrix transpose.
  • Tensor Operations: A mode-n fiber fixes every tensor index except the nth, producing a vector in the nth-mode dimension.For X ∈ R^(I_1×I_2×···×I_N), the mode-n fiber lies in R^I_n.
  • Tensor Operations: The mode-n product Y := X ×_n A left-multiplies every mode-n fiber of X by A.The resulting fiber is Y(i_1,…,:,…,i_N) = A X(i_1,…,:,…,i_N).
  • Regularization: Total-variation regularizers use forward finite differences to encode spatial, spatial–spectral, and spatial–temporal variation for matrices, hyperspectral images, and color videos.The operators include horizontal and vertical differences, with additional spectral or temporal differences for the corresponding tensor data.

A. General Subspace-Induced Tensor Orthogonal Split

TOSS lifts orthogonal projection from vectors to tensors by applying a prescribed subspace projector consistently to every fiber along one mode. The resulting components reconstruct the tensor exactly, remain fiber-wise orthogonal, preserve energy, and admit a reflection-based interpretation.

  • Definition and construction: TOSS applies the same orthogonal projector to every mode-n fiber, lifting vector-level subspace splitting into a coherent tensor decomposition.The prescribed mode is selected according to the organization of the data and the expected shared structure.
  • Orthogonality properties: The split places one component in a prescribed subspace and the other in its orthogonal complement, with orthogonal fibers for every fixed index tuple.This is the tensor analogue of x+ ∈ U and x− ∈ U⊥ with zero inner product.
  • Fundamental properties: X = X+ + X− gives TOSS the exact reconstruction property.The paper explicitly names this identity the exact reconstruction property of TOSS.
  • Operator interpretation: The reflection operator Rn leaves the subspace component unchanged and flips the sign of the orthogonal residual.Thus X+ and X− occupy the +1 and −1 eigenspaces, respectively, giving the split a geometric interpretation.
  • Fundamental properties: Because the components are orthogonal, their Frobenius energies add according to the tensor Pythagorean identity.The proof uses the vanishing Frobenius inner product ⟨X+, X−⟩ = 0.

B. Rank-One TOSS Induced by a Single Direction

Rank-one TOSS restricts the mode-n subspace to one unit direction µ, producing a separable dominant component and an orthogonal residual. All dominant mode-n fibers are collinear, while the split retains a geometric and energy-preserving interpretation.

  • Rank-one specialization: Rank-one TOSS uses a one-dimensional mode-n subspace spanned by a single dominant direction µ.The term rank-one refers to the subspace dimension and the resulting rank-one structure along mode n.
  • Fiber-wise split: Each mode-n fiber is decomposed into its projection onto µ and the orthogonal remainder.The projection coefficient is determined fiber-wise, so the construction applies consistently across the tensor.
  • Structural consequence: All mode-n fibers of X+ are collinear and lie in span{µ}, giving X+ a rank-one structure along mode n.This explicitly captures a common direction across the splitting mode.
  • Geometric interpretation: The associated involution is a Householder reflection, and the split remains symmetric under that operator.X+ captures fibers aligned with µ, whereas X− represents deviations from that direction.
  • Application-oriented interpretation: Rank-one TOSS separates consistent structure from orthogonal residuals such as noise, illumination changes, color fluctuations, or compression artifacts.The interpretation is motivated for color images and hyperspectral images with shared cross-channel or cross-band structure.

IV. APPLICATION EXAMPLES OF RANK-ONE TOSS

The paper applies rank-one TOSS to multidimensional data whose variation along a mode is dominated by consistent latent factors. Hyperspectral restoration and color-video background modeling serve as representative application tasks.

  • Application scope: TOSS is suited to data with consistent latent factors along one mode and residual, noisy, or localized variations in the complementary component.The paper identifies hyperspectral imaging, video analysis, medical imaging, and multi-sensor fusion as relevant settings.
  • Representative applications: Rank-one TOSS is applied to hyperspectral image restoration and color video background modeling as two representative validation tasks.The paper formulates optimization models and develops efficient algorithms for these applications.

A. Hyperspectral Image Restoration

For hyperspectral restoration, the observed image is modeled with dense and sparse corruption, while rank-one TOSS separates a dominant spectral component from orthogonal residual variations. PCA supplies a fixed dominant spectral direction for denoising, and spatial regularization is imposed on the coefficient map.

  • Observation model: The hyperspectral observation contains a clean image X, dense noise N, and sparse corruption S such as impulse noise, stripes, or dead pixels.The model explicitly distinguishes dense corruption from structured sparse artifacts.
  • Spectral decomposition: Rank-one TOSS splits the clean image along the spectral mode into a dominant component aligned with one spectral direction and an orthogonal residual.The spectral mode is mode-3 for Y ∈ R^H×W×B.
  • Separable representation: The dominant component has the separable form X+(:, :, b) = µ_b A, where A is the spatial coefficient map.The coefficient at each spatial location is a(i, j) = µ^T X(i, j, :).
  • Direction selection: For hyperspectral denoising, µ is obtained by normalizing the first principal component of the observed image and then held fixed during optimization.The observed image is reshaped along the spectral mode before PCA is applied.
  • Regularization: A spatial total-variation regularizer is imposed on A because it represents the main piecewise-smooth spatial content while preserving structural edges.The residual X− collects spectral-spatial information not represented by the dominant rank-one component.

B. Color Video Background Modeling

Rank-one TOSS models a color video as a temporally aligned background plus a residual containing foreground and other temporal variations. The formulation is further regularized to improve robustness when direct splitting is contaminated by complex motion.

  • Model formulation: Rank-one TOSS splits a color video into a background component aligned with one temporal direction and a residual containing foreground objects and temporal variations.The split is performed along mode 4.
  • Model formulation: The background component has a separable fourth-mode representation, with each spatial-color location sharing the temporal direction μ.Its slices satisfy X+(:, :, :, t) = μ_t A, where A is the spatial-color coefficient tensor.
  • Model formulation: The coefficient tensor A represents the spatial-color content associated with the common temporal direction.Projection coefficients are obtained by projecting each mode-4 fiber onto μ.
  • Robust modeling: Direct rank-one splitting can be satisfactory for simple sequences but may produce contaminated backgrounds under complex foreground motion or temporal fluctuations.The contamination motivates adding structural priors for more accurate and robust modeling.
  • Robust modeling: The resulting optimization model uses regularization parameters λ1, λ2, and λ3 together with spatial and spatiotemporal total-variation terms.The algorithm derivation is provided in the supplementary materials.

V. EXPERIMENTAL RESULTS

The experiments evaluate TOSS-based methods on hyperspectral restoration and color video background modeling. For color video, the temporal direction is chosen uniformly to represent the mean temporal background direction.

  • Experimental scope: The experimental section validates the proposed method on hyperspectral image restoration and color video background modeling.These are presented as two practical real-world applications.
  • Color video setting: Color video background modeling uses a uniform temporal vector μ to project onto the mean temporal direction shared across video frames.This choice captures the dominant consistent background component.
  • Experimental setting: Experiments run in MATLAB 2024b under Windows 11 on a personal computer with a 3.90GHz CPU and 64GB of memory.

A. Experiments on Hyperspectral Image Restoration

Hyperspectral restoration experiments test rank-one TOSS under pure Gaussian and mixed noise across three standardized datasets. The method is reported to achieve strong quantitative, visual, and computational performance, with gains attributed to orthogonal component splitting.

  • Experimental setup: The experiments use Pavia, DCMall, and RemoteImage, with every sample preprocessed to 200 × 200 × 80.
  • Experimental setup: Four noise cases combine Gaussian noise and salt-and-pepper noise, including pure Gaussian settings G = 0.1 and G = 0.2.Mixed cases use S = 0.1 with G = 0.01 or G = 0.1.
  • Restoration behavior: Rank-one TOSS separates dominant contours, boundaries, and layouts from orthogonal local details, then recombines both components for denoising while preserving structure.
  • Quantitative results: Across all three datasets and four metrics, the proposed method outperforms competing methods under pure Gaussian noise, including at G = 0.2.The metrics are PSNR, SSIM, FSIM, and ERGAS.
  • Quantitative results: 2.118 dB PSNR improvement and 19.83% ERGAS reduction occur on RemoteImage under S = 0.1 and G = 0.01 versus TCTV.
  • Visual results: Visual comparisons show sharper edges and more faithful local details, while TNN, LRTV, and TLR-SSTV exhibit color distortion under Gaussian noise.The proposed method is reported to preserve higher color fidelity.
  • Interpretation: The paper attributes the performance and efficiency advantage over TV-based baselines to orthogonal splitting, which lets TV act on complementary components without mutual interference.The split is also linked to smoother convergence and faster per-iteration updates.
  • Computational efficiency: 39.67 seconds is the proposed method’s competitive average runtime, compared with 55.38 s for SNN and 75.66 s for TNN.The reported runtime averages results across three HSI datasets and different noise cases.

B. Experiments on Color Video Background Modeling

Experiments on the SBMnet dataset evaluate color-video background modeling across diverse motion and camera conditions using visual and quantitative comparisons. The proposed method achieves the strongest overall results, while remaining relatively robust in challenging sequences.

  • Experimental setup: The study evaluates six SBMnet videos using AGE, PSNR, MSSSIM, and CQM to assess background-modeling quality.Higher PSNR, MSSSIM, and CQM are better, while lower AGE is better.
  • Visual results: Visual comparisons show favorable backgrounds across several sequences, although all methods exhibit moving-object shadows on “boulevard-Jam,” while ours remains relatively better.The method also remains robust on “badminton” and clearly outperforms several baselines on “BusStation.”
  • Quantitative results: Our method achieves the best performance across all four metrics on every video, including the lowest AGE and highest PSNR, MSSSIM, and CQM.It exceeds the second-best method by an average of approximately 3.1% in CQM across the six videos.

VI. CONCLUSION

TOSS provides a principled tensor framework for orthogonally separating subspace-aligned information from residual variations, with rank-one TOSS studied theoretically and applied to two restoration and modeling tasks. The conclusion identifies broader applications, richer subspaces, adaptive direction estimation, and deep-learning integration as future directions.

  • VI. CONCLUSION: TOSS splits high-dimensional tensors into orthogonal components along a prescribed mode under an underlying subspace structure.It extends orthogonal projection from vector spaces to tensors in a consistent, fiber-wise manner.
  • VI. CONCLUSION: Rank-one TOSS was investigated as an important special case and applied to hyperspectral image restoration and color video background modeling.Corresponding optimization models and efficient ADMM-based algorithms were developed for both tasks.
  • VI. CONCLUSION: Experimental results demonstrated the effectiveness of the proposed methods and validated rank-one TOSS in the representative applications.
  • VI. CONCLUSION: The broader potential of general TOSS remains largely unexplored beyond the two validated representative tasks.Future work includes richer subspace configurations and evaluation across a wider range of multidimensional data-analysis problems.
  • VI. CONCLUSION: Current models rely on a prescribed dominant direction, motivating more adaptive, data-driven estimation of X+ and X−.
  • VI. CONCLUSION: TOSS also offers a potential interface with deep learning through cooperative architectures for subspace-consistent and orthogonal residual components.The proposed direction seeks to combine the theory’s interpretability with modern data-driven models’ expressive power.
Loading 2608.29638v1…