Abstract:In real-world scenarios, complex data such as multispectral images and multi-frame videos inherently exhibit robust low-rank property. This property is vital for multi-dimensional inverse problems, such as tensor completion, spectral imaging reconstruction, and multispectral image denoising. Existing tensor singular value decomposition (t-SVD) definitions rely on hand-designed or pre-given transforms, which lack flexibility for defining tensor nuclear norm (TNN). The TNN-regularized optimization problem is solved by the singular value thresholding (SVT) operator, which leverages the t-SVD framework to obtain the low-rank tensor. However, it is quite complicated to introduce SVT into deep neural networks due to the numerical instability problem in solving the derivatives of the eigenvectors. In this paper, we introduce a novel data-driven generative low-rank t-SVD model based on the learnable orthogonal transform, which can be naturally solved under its representation. Prompted by the linear algebra theorem of the Householder transformation, our learnable orthogonal transform is achieved by constructing an endogenously orthogonal matrix adaptable to neural networks, optimizing it as arbitrary orthogonal matrices. Additionally, we propose a low-rank solver as a generalization of SVT, which utilizes an efficient representation of generative networks to obtain low-rank structures. Extensive experiments highlight its significant restoration enhancements.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the low - rank tensor representation and recovery in multi - dimensional inverse problems. Specifically, the author focuses on how to effectively capture and utilize the low - rank characteristics in complex multi - dimensional data (such as hyperspectral images, multi - frame videos, etc.) to solve problems such as tensor completion, spectral imaging reconstruction, and hyperspectral image denoising. ### Main problems 1. **Limitations of existing methods**: - The existing tensor singular value decomposition (t - SVD) depends on predefined or given transformations and lacks flexibility. - The tensor nuclear norm (TNN) regularization optimization problem is solved by the singular value thresholding (SVT) operator, but there is a numerical instability problem when introducing SVT in deep neural networks. 2. **Data adaptability and theoretical guarantees**: - Current methods are not flexible enough when dealing with different data instances and are difficult to adapt to different application scenarios. - A method that can both maintain theoretical guarantees and adapt to data changes is required. ### Solutions To overcome the above problems, the author proposes a new generative low - rank t - SVD model based on learnable orthogonal transformation (OTLRM). The main features of this model are as follows: 1. **Learnable orthogonal transformation**: - Use Householder transformation to construct endogenous orthogonal matrices to ensure the orthogonality and data adaptability of the transformation. - Optimize the parameter matrix \( W \) through a neural network to achieve adaptive adjustment for each data set. 2. **Generative framework**: - Under the t - SVD representation, OTLRM can directly generate the desired tensor and guide the observed data in the DNN optimization framework. - Introduce a dense rank estimation operator to store and enrich the rank information of each frequency band in the transform domain. 3. **Numerical stability**: - Use the gradient descent algorithm to solve the low - rank problem in the DNN optimization framework, avoiding the numerical instability problem caused by SVD. ### Application scenarios This model has been verified in the following tasks: - **Tensor completion**: Recover complete multi - dimensional data from partially observed data. - **MSI reconstruction in snapshot compressive imaging system (CASSI)**: Recover the original hyperspectral image from the encoded measurement. - **MSI denoising**: Recover a clean hyperspectral image from a noise - contaminated observation. ### Experimental results Experiments show that OTLRM significantly outperforms existing methods on multiple data sets, especially in the case of low sampling rates. By introducing the orthogonal total variation (OTV) loss, the generative ability and image quality are further improved. In conclusion, this paper aims to solve the low - rank tensor representation and recovery problems in multi - dimensional inverse problems by introducing learnable orthogonal transformations and a generative framework, thereby improving the accuracy and robustness of data recovery.

OTLRM: Orthogonal Learning-based Low-Rank Metric for Multi-Dimensional Inverse Problems

Online Robust Low-Rank Tensor Modeling for Streaming Data Analysis

Multilinear Low-Rank Tensors on Graphs & Applications

Learnable Spatial-Spectral Transform-Based Tensor Nuclear Norm for Multi-Dimensional Visual Data Recovery

A Scale-Invariant Relaxation in Low-Rank Tensor Recovery with an Application to Tensor Completion

T$^2$LR-Net: An Unrolling Reconstruction Network Learning Transformed Tensor Low-Rank prior for Dynamic MR Imaging

Low-Rank Tensor Learning by Generalized Nonconvex Regularization

Low-rank Tensor Learning with Nonconvex Overlapped Nuclear Norm Regularization

A novel minimization framework of tensor tubal rank and its multi-dimensional image completion application

T-SVD Based Non-convex Tensor Completion and Robust Principal Component Analysis.

Sparse and Low-Rank High-Order Tensor Regression via Parallel Proximal Method

On The Tensor Svd And The Optimal Low Rank Orthogonal Approximation Of Tensors

Adaptive total variation and second-order total variation-based model for low-rank tensor completion

Learnable Transform-Assisted Tensor Decomposition for Spatio-Irregular Multidimensional Data Recovery

Tensor Q-rank: New Data Dependent Definition of Tensor Rank

Low-Rank Tensor Function Representation for Multi-Dimensional Data Recovery

Operator SVD with Neural Networks via Nested Low-Rank Approximation

Handling The Non-Smooth Challenge in Tensor SVD: A Multi-Objective Tensor Recovery Framework

Tensor completion via joint reweighted tensor Q -nuclear norm for visual data recovery

Robust Low-Rank Tensor Recovery with Rectification and Alignment

Nonlinear Transform Induced Tensor Nuclear Norm for Tensor Completion