Article
Environmental Sciences
Andong Wang, Guoxu Zhou, Qibin Zhao
Summary: This study conducts a rigorous analysis for the problem of robust tensor completion, proposing a new estimator based on *L-SVD framework and ADMM algorithm to effectively recover the unknown three-way tensor, showing good optimization performance and statistical properties.
Article
Geochemistry & Geophysics
Mingwen Shao, Chao Wang, Wangmeng Zuo, Deyu Meng
Summary: This article proposes a novel generative adversarial network-based unified framework for missing remote sensing (RS) image reconstruction. The framework is capable of various reconstruction tasks given only single-source data as input, and extensive experiments demonstrate its uncompromising performance on diverse missing restoration.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Environmental Sciences
Zhihong Chen, Peng Zhang, Yu Zhang, Xunpeng Xu, Luyan Ji, Hairong Tang
Summary: This paper proposes a new frequency spectrum-modulated tensor completion method (FMTC) for the reconstruction of multi-temporal remote sensing images. By rearranging and Fourier transforming the remote sensing images, the joint low-rank spatial-temporal constraint is achieved using the frequency spectrum. The simulated and real data experiments demonstrate the applicability and stability of FMTC for different land-cover types and missing sizes. Compared with other algorithms, FMTC performs better in terms of spectral accuracy and temporal continuity.
Article
Geochemistry & Geophysics
Mingwen Shao, Chao Wang, Tianjun Wu, Deyu Meng, Jiancheng Luo
Summary: This letter proposes a novel generative adversarial network-based method for missing data reconstruction in remote sensing images, capable of various reconstruction tasks with only single source data as input. The proposed model, incorporating special convolutions and attention mechanism, along with perceptual and multiscale adversarial losses, demonstrates outstanding performance in generating coherent structures with better details.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
(2022)
Article
Computer Science, Hardware & Architecture
Wenjian Ding, Zhe Sun, Xingxing Wu, Zhenglu Yang, Jordi Sole-Casals, Cesar F. Caiafa
Summary: This study addresses the issue of missing data in multi-channel audio signals using tensor completion algorithms and demonstrates their superior predictive performance in signal restoration.
COMPUTERS & ELECTRICAL ENGINEERING
(2022)
Article
Geochemistry & Geophysics
Xuemin Zhang, Jianwei Ma, Siwei Yu
Summary: Multidimensional prestack seismic data reconstruction can be seen as a low-rank tensor completion problem. To address the suboptimal solution issue caused by the nuclear norm-based relaxation, we propose a nonconvex logDet function as a smooth approximation for the tensor rank. By applying an alternating direction method of multipliers (ADMMs) algorithm, we achieve remarkable reconstruction performance on synthetic and land data survey.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2023)
Article
Computer Science, Information Systems
Chuan Chen, Zhe-Bin Wu, Zi-Tai Chen, Zi-Bin Zheng, Xiong-Jun Zhang
Summary: This paper studies the Robust Tensor Completion (RTC) problem, which combines Low-Rank Tensor Completion (LRTC) and Robust Principal Component Analysis (RPCA) to recover low rank components and separate sparse components from tensor data, addressing signal corruptions and missing values. The proposed model utilizes Tensor-Train rank (TT rank), auto-weighted mechanism, and Tree Ket Augmentation (Tree-KA) to achieve effective tensor completion and augmentation. Extensive numerical experiments demonstrate the effectiveness of the model compared to other state-of-the-art methods.
INFORMATION SCIENCES
(2021)
Article
Computer Science, Interdisciplinary Applications
Liyue Shen, Wentao Zhu, Xiaosong Wang, Lei Xing, John M. Pauly, Baris Turkbey, Stephanie Anne Harmon, Thomas Hogue Sanford, Sherif Mehralivand, Peter L. Choyke, Bradford J. Wood, Daguang Xu
Summary: A novel multi-domain image completion method using GAN is proposed to extract shared content encoding and separate style encoding across multiple domains. The learned representation in multi-domain image completion can be leveraged for high-level tasks, like segmentation, by introducing a unified framework consisting of image completion and segmentation.
IEEE TRANSACTIONS ON MEDICAL IMAGING
(2021)
Article
Computer Science, Artificial Intelligence
Qian Zhao, Yuji Lin, Fengxingyu Wang, Deyu Meng
Summary: Weighted nuclear norm characterizes the low-rank structure of a matrix and has been successfully used in matrix completion. However, the fixed weighting functions used in previous studies may not accurately capture the underlying structure of the data matrix, especially in dynamic estimation processes. To address this issue, we propose an adaptive weighting function strategy for low-rank matrix/tensor completion. Our empirical studies demonstrate the effectiveness of this approach compared to representative methods.
INTERNATIONAL JOURNAL OF MACHINE LEARNING AND CYBERNETICS
(2023)
Article
Automation & Control Systems
Yuan Gao, Laurence T. Yang, Jing Yang, Dehua Zheng, Yaliang Zhao
Summary: In this article, a jointly low-rank tensor completion method is proposed for logistics data completion, which constructs multiple periodic subtensors by setting an appropriate time window and performs jointly low-rank completion and imputation. Experimental results demonstrate the promising performance of the proposed method compared with other state-of-the-art competitors in logistics systems.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS
(2023)
Article
Geochemistry & Geophysics
Yanxin Xi, Luyan Ji, Weitun Yang, Xiurui Geng, Yongchao Zhao
Summary: This article proposes four multi-target detection algorithms for multitemporal remote sensing data by combining filter tensor analysis with multiple target constraints, fully exploiting the time-series information. The experimental results show the effectiveness and superiority of the proposed methods.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Geochemistry & Geophysics
Dawei Liu, Mauricio D. Sacchi, Wenchao Chen
Summary: The study on five-dimensional seismic reconstruction is gaining attention, and this paper proposes two efficient methods to utilize the low rank structure of tensors and reduce computational costs.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Geochemistry & Geophysics
Mu Xia, Kun Jia
Summary: An improved method based on multitemporal dictionary learning is proposed for reconstructing remote sensing data contaminated by large and thick clouds. The method initializes the contaminated target image using adjacent cloud-free reference images, produces reconstructed images from each reference image using dictionary learning and sparse representation methods, and combines them with original uncontaminated pixels to generate the final result. Visual and quantitative analyses show that the proposed method outperforms commonly used methods in accurately and effectively reconstructing data contaminated by large and thick clouds.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Engineering, Electrical & Electronic
Teng-Yu Ji, Xi-Le Zhao, Dong-Lin Sun
Summary: Existing low-rank tensor completion methods have limitations in considering the low-rank characteristic, which is not applicable to high-rank real-world data. Therefore, this research proposes a kernel low-rank tensor completion model, which maps high-rank data into a low-rank feature space through kernel mapping to better utilize the global correlation relationship for estimating the missing information.
IEEE SIGNAL PROCESSING LETTERS
(2022)
Article
Computer Science, Information Systems
Rui Xu, Xiaoyang Ma, Runze Zhou, Jinshuai Zhao, Ying Wang
Summary: This paper proposes a novel approach for completing missing harmonic data based on a data-driven method. By constructing an adjacency matrix and unconstrained optimization function, the method effectively reduces data dependence and does not require optimization strategy of phasor measurement units. The augmented Lagrange iteration algorithm is used to iteratively solve the unconstrained optimization function, demonstrating effectiveness and applicability through artificial and field data verification.
Article
Mathematics, Applied
Michael K. Ng, Andy Yip
Summary: This paper analyzes two-layer Graph Convolution Networks (GCNs), focusing on their generalization and stability. The study also explores the impact of data scaling on the network's stability.
ANALYSIS AND APPLICATIONS
(2023)
Article
Geochemistry & Geophysics
Lina Zhuang, Michael K. Ng, Yao Liu
Summary: This article introduces the working principle of a hyperspectral pushbroom sensor and addresses the cross-track illumination error issue. A mathematical model is developed to describe the image formation process corrupted by this error and additive Gaussian noise. A new method called HyCIC is proposed, which corrects the cross-track illumination using column mean compensation and attenuates the Gaussian noise using low-rank constraint.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2023)
Article
Mathematics, Applied
Wei-Wei Xu, Michael K. Ng
Summary: A High-Order Generalized Singular Value Decomposition (HO-GSVD) is used to compare multiple matrices {Ai} (N) (i=1) with different row dimensions by examining their generalized singular values {sigma i, k} Ni=1. The significance of the k-th basis vector on the right hand side of the matrix from HO-GSVD for multiple matrices {A(i)} (N) (i=1) can be determined by the ratio values of sigma(i, k)/sigma(j, k). This paper proposes and studies a new matrix maximization model for computing these ratio values from A(1), ..., A(N), which can be solved using the Newton method on Lie Groups with well-defined initial values. Numerical examples are provided to demonstrate the fast performance of the proposed method in solving the optimization model compared to existing algorithms and the Riemannian Newton method.
JOURNAL OF SCIENTIFIC COMPUTING
(2023)
Article
Engineering, Electrical & Electronic
Huan Huan Zhang, He Ming Yao, Lijun Jiang, Michael Ng
Summary: This letter proposes a novel deep learning-based fast solver for the electromagnetic forward process. The solver is based on a deep conditional convolutional autoencoder (DCCAE) consisting of a complex-valued deep convolutional encoder network and its corresponding decoder network. The proposed solver can accurately predict the electromagnetic field of a target domain in real-time applications, significantly reducing computation time compared to conventional methods.
IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS
(2023)
Article
Mathematics, Applied
Guang-Jing Song, Xue-Zhong Wang, Michael K. Ng
Summary: This paper investigates the low rank third-order tensor completion problems and proposes a solution using Riemannian optimization methods. With suitable incoherence conditions, the proposed method can converge to the underlying low rank tensor with high probability and the required number of sampling entries for convergence is derived. Numerical experiments demonstrate that the proposed method outperforms other methods in terms of computational time and number of sampling entries.
JOURNAL OF COMPUTATIONAL AND APPLIED MATHEMATICS
(2023)
Article
Computer Science, Artificial Intelligence
Ye Liu, Junjun Pan, Michael K. Ng
Summary: This paper proposes a deep neural network called the Tucker network derived from the Tucker format and analyzes its expressive power. The results show that the Tucker network has exponentially higher expressive power than the shallow network. Additionally, the paper discusses the expressive power between the hierarchical Tucker tensor network (HT network) and the proposed Tucker network. Experimental results validate the theoretical findings and demonstrate the superior performance of the Tucker network and deep Tucker network compared to the shallow network and HT network on three datasets.
Article
Computer Science, Information Systems
Junren Chen, Cheng-Long Wang, Michael K. Ng, Di Wang
Summary: In this paper, a uniformly dithered 1-bit quantization scheme for high-dimensional statistical estimation is proposed. The scheme includes truncation, dithering, and quantization as typical steps. The scheme is applied to the estimation problems of sparse covariance matrix estimation, sparse linear regression, and matrix completion. New estimators based on 1-bit quantized data are proposed. The rates of the estimators achieve minimax rates up to logarithmic factors in the sub-Gaussian regime, and improve existing results in the heavy-tailed regime. The approach to 1-bit matrix completion is robust to pre-quantization noise with unknown distribution.
IEEE TRANSACTIONS ON INFORMATION THEORY
(2023)
Article
Computer Science, Information Systems
Junren Chen, Michael K. Ng
Summary: This paper focuses on the signal recovery problem in phase-only compressive sensing. It addresses two open questions regarding the uniform recovery guarantee and exact recovery of complex signal. The authors prove that all complex sparse signals or low-rank matrices can be uniformly, exactly recovered using a near optimal number of complex Gaussian measurement phases. They also propose methods for handling complex signals and demonstrate the stability of uniform recovery under bounded noise. Experimental results confirm the validity of the theoretical findings.
IEEE TRANSACTIONS ON INFORMATION THEORY
(2023)
Article
Computer Science, Artificial Intelligence
Guangjing Song, Michael K. Ng, Tai-Xiang Jiang
Summary: In this article, a new alternating projection method is developed to compute nonnegative low-rank matrix approximation for nonnegative matrices. The proposed method reduces computational cost by approximating the projection onto the manifold using the tangent space of the point in the manifold. Numerical examples demonstrate that the proposed method outperforms nonnegative matrix factorization methods in terms of computational time and accuracy.
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING
(2023)
Article
Mathematics, Applied
Junren Chen, Michael K. Ng
Summary: This paper investigates the phase-only reconstruction problem, focusing on recovering a complex-valued signal x in Cd from the phase of Ax. Uniqueness conditions are derived using discriminant matrices, determining if the signal can be uniquely reconstructed. The minimum measurement number is also examined, with at least 2d but no more than 4d-2 measurements needed for reconstruction of all x∈Cd. Practical and general uniqueness conditions are provided for the phase-only reconstruction in Rd, and the results can be extended to affine phase-only reconstruction where the phase of Ax + b is observed for some b∈Cm.
SIAM JOURNAL ON APPLIED MATHEMATICS
(2023)
Article
Mathematics, Applied
Junjun Pan, Michael K. Ng
Summary: This paper introduces the nonnegative matrix factorization model (NMF) and its extended form, coseparable NMF (CoS-NMF), and studies their mathematical properties and relationships with other matrix factorization methods. The paper also proposes an optimization method for CoS-NMF and verifies its effectiveness and superiority through experiments.
SIAM JOURNAL ON MATRIX ANALYSIS AND APPLICATIONS
(2023)
Article
Computer Science, Artificial Intelligence
Lina Zhuang, Michael K. Ng, Lianru Gao, Joseph Michalski, Zhicheng Wang
Summary: The performance of deep learning-based denoisers is highly dependent on the quantity and quality of the training data. However, in hyperspectral remote sensing areas, paired noisy-clean training images are generally unavailable. To overcome this, this work uses self-supervised learning to train a model that can learn one part of noisy input from another part. The proposed Eigenimage2Eigenimage (E2E) framework converts the HSI denoising problem into an eigenimage denoising problem and generates noisy-noisy paired training eigenimages. Experimental results demonstrate that the proposed method outperforms other existing deep learning methods for denoising HSIs. A MATLAB demo is provided for reproducibility.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS
(2023)
Article
Mathematics, Applied
Yun-Yang Liu, Xi-Le Zhao, Guang-Jing Song, Yu-Bang Zheng, Michael K. Ng, Ting-Zhu Huang
Summary: Motivated by the success of fully-connected tensor network (FCTN) decomposition, this study proposes two FCTN-based models for the robust tensor completion (RTC) problem. The first model, named RNC-FCTN, directly applies FCTN decomposition for the RTC problem. An algorithm based on proximal alternating minimization (PAM) is developed to solve RNC-FCTN. The second model, named RC-FCTN, uses the FCTN nuclear norm as a convex surrogate function and applies robust convex optimization for RTC. An algorithm based on alternating direction method of multipliers (ADMM) is developed for RC-FCTN.
INVERSE PROBLEMS AND IMAGING
(2023)
Article
Mathematics, Applied
Huanmin Ge, Wengu Chen, Michael K. Ng
Summary: In this paper, we propose a novel model, the weighted l(1)/l(2) minimization, which incorporates partial support information into the standard l(1)/l(2) minimization to recover sparse signals from linear measurements. We establish the restricted isometry property based conditions for sparse signal recovery using the weighted l(1)/l(2) minimization in both noiseless and noisy cases. Our results show that these conditions are weaker than the analogous conditions for standard l(1)/l(2) minimization when the accuracy of the partial support information is at least 50%. Additionally, we develop effective algorithms and validate our results through extensive numerical experiments using synthetic data in both noiseless and noisy cases.
INFORMATION AND INFERENCE-A JOURNAL OF THE IMA
(2023)
Article
Computer Science, Artificial Intelligence
Hanrui Wu, Yuguang Yan, Michael Kwok-Po Ng
Summary: In this paper, a novel model called Hypergraph Collaborative Network (HCoN) is proposed, which considers the information from both previous vertices and hyperedges to achieve informative latent representations and introduces the hypergraph reconstruction error as a regularizer to learn an effective classifier. Experimental results demonstrate that the proposed method outperforms the baseline methods.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
(2023)
