Abstract
Tensor sparse modeling as a promising approach, in the whole of science and engineering has been a huge success. As is known to all, various data in practical application are often generated by multiple factors, so the use of tensors to represent the data containing the internal structure of multiple factors came into being. However, different from the matrix case, constructing reasonable sparse measure of tensor is a relatively difficult and very important task. Therefore, in this paper, we propose a new tensor sparsity measure called Tensor Full Feature Measure (FFM). It can simultaneously describe the feature information of each dimension of the tensor and the related features between two dimensions, and connect the Tucker rank with the tensor tube rank. This measurement method can describe the sparse features of the tensor more comprehensively. On this basis, we establish its non-convex relaxation, and apply FFM to low rank tensor completion (LRTC) and tensor robust principal component analysis (TRPCA). LRTC and TRPCA models based on FFM are proposed, and two efficient Alternating Direction Multiplier Method (ADMM) algorithms are developed to solve the proposed model. A variety of real numerical experiments substantiate the superiority of the proposed methods beyond state-of-the-arts.
Introduction
In fields like medical imaging, remote sensing, and video processing, data is naturally represented as tensors (multi-dimensional arrays). A core challenge in "tensor recovery"—tasks like filling in missing pixels (completion) or removing noise (robust PCA)—is defining what "sparsity" or "low-rankness" actually means for a tensor.
Unlike matrices, where the rank is well-defined, tensors have multiple competing rank definitions (e.g., CP rank, Tucker rank, Tubal rank). Existing measures often fall into two traps: oversimplification, where dimensions are treated independently, or specificity, where methods focus too heavily on specific dimensional pairings. This paper introduces the Tensor Full Feature Measure (FFM) to bridge this gap by capturing both unique features of individual dimensions and complex correlations between pairs.
Methodology
The proposed Tensor Full Feature Measure (FFM) is defined as a weighted sum of ranks across two types of "unfoldings":
- Mode-$n$ unfoldings: Captures features inherent to a single dimension (e.g., spatial continuity).
- Mode-$k_1, k_2$ unfoldings: Captures the correlation between any two different dimensions (e.g., how a specific spatial feature evolves over time).
Key Innovations
- Non-convex Relaxation: Instead of the standard Nuclear Norm, the authors use a logarithmic sum function ($\sum \log(\sigma + \epsilon)$). This non-convex approach provides a much "tighter" approximation of the actual rank, leading to higher recovery accuracy.
- Unified Framework: FFM is the first measure to explicitly relate Tucker rank and tensor tube rank within a single optimization objective.
- ADMM Solver: To handle the complex optimization, the authors developed two algorithms based on the Alternating Direction Method of Multipliers (ADMM)—FFMTC for Tensor Completion and FFMTRPCA for Robust PCA.
Experimental Results
The FFM-based models were evaluated against 12 state-of-the-art competitors across several real-world datasets, including Multispectral Images (MSI), MRI scans, and hyperspectral videos.
Quantitative Performance
As shown in the table below, the FFMTC method consistently outperforms existing methods across all metrics (PSNR, SSIM, FSIM, and ERGAS). For instance, on MSI data with only 5% of pixels available, FFMTC achieved a PSNR of 40.350 dB, outperforming the next best method (NWSTNN) by roughly 3 dB.
Qualitative Visual Fidelity
Qualitative results demonstrate that FFM preserves fine textures and edges that other methods blur or fail to recover. This is particularly evident in complex scenes where structural correlations are vital for accurate reconstruction.
Conclusion
The Tensor Full Feature Measure (FFM) offers a comprehensive approach to tensor recovery by integrating Tucker-style and Tubal-style ranks. By utilizing a non-convex log-sum penalty and an efficient ADMM solver, the proposed methods achieve superior performance in both quantitative metrics and visual fidelity. While computationally heavier than simpler methods, the significant gains in accuracy make FFM a powerful tool for high-dimensional data restoration.
Summary updated