Georgia Institute of Technology
Accelerated Tensor Robust Algorithms for Hyperspectral Imaging and Video Processing
Abstract
dc:description.abstractIn recent years, the application of tensor-based methods to high-dimensional data has gained considerable attention, particularly for tasks involving denoising, classification, and compression of complex data structures such as hyperspectral images. This thesis presents novel approaches to enhance Tensor Robust Principal Component Analysis (TRPCA), addressing challenges such as computational efficiency, noise removal, and real-time processing. Firstly, the thesis presents a new online robust principal component analysis (RPCA) algorithm that recursively decomposes incoming data into low-rank and sparse components. Unlike traditional approaches that operate on data vectors, this method preserves the multi-dimensional structure of data, such as video frames. It is based on the recently proposed tensor singular value decomposition (T-SVD) and incorporates a convex optimization-based approach for recovering the sparse component and updating the low-rank component using incremental T-SVD. An efficient tensor convolutional extension to the Fast Iterative Shrinkage Thresholding Algorithm (FISTA) is also proposed, significantly speeding up the optimization process. The effectiveness of this online tensor-RPCA is demonstrated through its application in background-foreground separation in video streams, where the foreground is modeled as a sparse signal and the background as a gradually changing low-rank subspace. Extensive experiments on real-world videos showcase the robustness and effectiveness of the proposed algorithm. Secondly, the thesis proposes a randomized blocked algorithm for tensor singular-value thresholding (T-SVT), aimed at reducing the computational demands of TRPCA when applied to noisy hyperspectral images. TRPCA has been successfully employed to reduce noise by employing a minimization involving a tensor nuclear norm and a \ell1-norm to separate the low-rank hyperspectral image from the sparse noise. However, the high computational complexity of TRPCA is primarily due to the implementation of the T-SVT operator, which typically involves performing full tensor singular value decomposition (T-SVD) followed by shrinking the singular values of the frontal slices in the frequency domain. The proposed randomized blocked algorithm incrementally finds the singular values until they fall below the threshold, leveraging compression achieved by the fast Fourier transform (FFT) to accelerate TRPCA significantly. Numerical experiments indicate that this method is much faster than traditional TRPCA approaches while maintaining classification accuracy. Finally, the tensor-robust CUR (TRCUR) method is introduced for hyperspectral data compression and denoising. This method heavily downsamples the input hyperspectral image to form small subtensors and performs TRPCA on these subtensors. The desired hyperspectral image is recovered by combining the low-rank solution of the subtensors using tensor CUR reconstruction. We provide theoretical guarantees showing that the desired low-rank tensor can be exactly recovered using our proposed TRCUR method. Numerical experiments demonstrate that our method is up to 14 times faster than performing TRPCA on the original input data, while maintaining the classification accuracy.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Mohammadpour Salut, Mohammad
- Advisor dc:contributor.advisor
-
- Anderson, David V.
- Committee members dc:contributor.committeemember
-
- Calhoun, Vince
- Davenport, Mark A.
- Romberg, Justin
- Xie, Yao
Subjects
dc:subject × 4Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/76946