Abstract
High-dimensional data recovery often relies on the Tensor Nuclear Norm (TNN) via the Tensor Singular Value Decomposition (t-SVD). While traditionally implemented using the Discrete Fourier Transform (DFT), this paper proposes a Framelet-based Tensor Nuclear Norm (F-TNN) that leverages the redundancy of tight wavelet frames. This approach induces a lower-rank representation in the transform domain, enhancing the recovery of missing or corrupted data. Experiments on MRI, multispectral, and video datasets show that F-TNN consistently outperforms state-of-the-art methods in both tensor completion and robust principal component analysis tasks.
Introduction
Third-order tensors are fundamental for representing multidimensional data such as MRI volumes, multispectral images (MSI), and video sequences. However, real-world data acquisition often results in incomplete or corrupted observations. Tensor completion aims to reconstruct the underlying low-rank tensor from these limited samples. The standard approach involves minimizing the Tensor Nuclear Norm (TNN), typically defined through the t-SVD which utilizes the Discrete Fourier Transform (DFT) along the third dimension. While effective, the DFT is a unitary transform that may not provide the sparsest possible representation for complex spatial-temporal correlations found in real-world data.
Methodology: The F-TNN Framework
The authors introduce the Framelet Representation of Tensor Nuclear Norm (F-TNN) to address the limitations of invertible transforms like the DFT or Discrete Cosine Transform (DCT).
Framelet-Based Sparsity
The core innovation is replacing the DFT with a tight wavelet frame (framelet) transform, specifically a piece-wise cubic B-spline system. Unlike the DFT, framelets are redundant. This redundancy is strategically used to achieve a sparser representation of the tensor's "tubes" (the third dimension).
As shown in the figure above, the framelet domain exhibits a significantly higher proportion of singular values near zero compared to the DFT or DCT domains. This lower "multi-rank" structure is critical for high-fidelity recovery from sparse data, as it captures spatial-shifting correlations more effectively.
Optimization via ADMM
The F-TNN is defined as the sum of the nuclear norms of the frontal slices of the framelet-transformed tensor. To solve the resulting convex optimization problems for tensor completion and Tensor Robust Principal Component Analysis (TRPCA), the authors employ the Alternating Direction Method of Multipliers (ADMM). The non-invertibility of the redundant framelet transform is managed using the Unitary Extension Principle (UEP), ensuring a stable and efficient reconstruction process.
Experimental Results
The F-TNN method was evaluated against several state-of-the-art techniques, including LRMC, HaLRTC, TMac, TNN, PSTNN, and DCTNN, across MRI, MSI, and video datasets.
Quantitative and Qualitative Performance
F-TNN consistently achieved superior results in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Feature Similarity (FSIM). For instance, in MRI completion at a 10% sampling rate, F-TNN reached a PSNR of 26.104, significantly outperforming the standard TNN (21.855) and DCT-based methods.
Visual inspections confirm that F-TNN excels at preserving fine details and sharp edges. In the MRI recovery task (see figure above), F-TNN maintains the structural integrity of the brain tissue even at extremely low sampling rates (SR = 0.1), where competing methods produce significant artifacts or overly smoothed results.
Conclusion
The F-TNN framework demonstrates that redundancy in transform domains can be a powerful asset for tensor recovery. By inducing a lower-rank structure through framelets, the method provides a more robust and accurate solution for tensor completion and denoising. While the redundant transform increases computational complexity, the gains in recovery quality make it a compelling choice for high-stakes applications like medical imaging and satellite surveillance.
Summary updated