Dissertation Announcement for Long Tian — 03/06/2019 at 2:00 PM
January 28, 2019
Dear Faculty, Graduate and Undergraduate Students,
You are cordially invited to my Ph. D. dissertation defense.
Dissertation Title: Low-rank and Sparse Decomposition for Hyperspectral Image Enhancement and Clustering
When: Wednesday, March 6, 2019, 2:00 pm
Where: Simrall 228
Candidate: Long Tian
Degree: Ph.D., Electrical and Computer Engineering
Committee:
Dr. Jenny Q. Du
Professor of Electrical and Computer Engineering
(Major Professor)
Dr. James E Fowler
Professor of Electrical and Computer Engineering
(Committee Member)
Dr. Nicolas H. Younan
Professor and Department Head of Electrical and Computer Engineering
(Committee Member)
Dr. Qingmin Meng
Professor of Geosciences
(Committee Member)
Abstract
In this dissertation new algorithms are developed to improve the performance of hyperspectral image analysis.
First, tensor based extended low-rank and sparse decomposition (TELRSD) is proposed for hyperspectral image classification using low-rank tensor part and hyperspectral detection using sparse tensor part. With this tensor based method, a hyperspectral image is truly processed in 3D data format rather than a converted 2D matrix, and information between spectral bands and spatial pixels maintain integrated during decomposition process. This proposed algorithm is compared with other state-of-the-art methods, and the experiment results show that TELRSD offers the best performance among all comparative algorithms.
Second, hyperspectral image clustering is investigated, which is an unsupervised task, aiming to group pixels into different groups without labeled information. The spatial-spectral based multi-view low-rank sparse subspace clustering (SSMLC) algorithm is proposed in this dissertation, which extended Low-rank sparse subspace clustering (LRSSC) with multi-view learning technique. In this algorithm, spectral and spatial views are created to generate multi-view dataset of a hyperspectral image, where spectral partition, morphological component analysis and principle component analysis are applied to create others views. Furthermore, kernel version of SSMLC (k-SSMLC) is also developed.
Finally, spectral clustering is applied for hyperspectral image clustering, which has been proved equivalent to less costly non-negative matrix factorization (NMF)-based clustering. In order to include local and nonlinear features in data source, orthogonal NMF (ONMF) and kernel NMF (kNMF) are proposed for better clustering performance. The non-linear orthogonal NMF (kONMF) combines both kernel and orthogonal constraints in NMF, which pushes up the clustering performance further. In this dissertation, multi-view kOGNMF is also developed for subspace clustering, where graph constraints and multi-view learning are used to fully utilize the original data information.
Best Regards,
Long Tian