Charilaos Kanatsoulis presented a webinar, "Hyperspectral Super-resolution: A Tensor Factorization Approach," on Friday, July 24 at 10:00 am EDT (UTC -4:00). The edited video is available through YouTube, and download as a .mp4 [112MB].
Charilaos is a Ph.D. candidate in the Department of Electrical and Computer Engineering (ECE) at the University of Minnesota (UMN)- Twin Cities, USA.
Abstract.
Diagram of the information in a hyperspectral image (HSI) and multispectral image (MSI) combined to produce a higher-resolution super-resolution image (SRI) of high spatial and spectral resolution.
Hyperspectral super-resolution refers (HSR) to the problem of producing a super-resolution image (SRI) of high spatial and spectral resolution. The task is very well-motivated, since super-resolution images are of great interest to multiple applications in agricultural analytics and remote sensing. However, hardware limitations of multiband sensors render the problem significantly costly and challenging. A natural way to overcome this issue is by fusing a hyperspectral image (HSI) and a multispectral image (MSI) to produce an SRI that admits fine spatial and spectral resolutions. This inverse problem remains ill-posed, since the number of pixels in hyperspectral and multispectral images combined, are fewer that those required for super-resolution. In this talk, the super-resolution problem is tackled from a tensor perspective. The multidimensional structure of the HSI and MSI is utilized to propose a coupled tensor factorization framework that is guaranteed to recover the SRI. Furthermore, this tensor approach operates with relaxed modelling assumptions compared to what was previously considered state-of-the-art. The proposed tensor framework works remarkably well under various and challenging scenarios and provides a new tool in the powerful multi-band imaging toolbox.
Supplementary/Reading materials:
IEEE Transactions on Signal Processing paper (arXiv version).
ICIP paper (RG).
background: Yokoya et al. review.
Presenter Biography.
Charilaos Kanatsoulis, speaker for webinar #56.
Charilaos Kanatsoulis is a Ph.D. candidate in the Department of Electrical and Computer Engineering (ECE) at the University of Minnesota (UMN), Twin Cities. He received his Diploma in electrical and computer engineering from the National Technical University of Athens, Greece, in 2014. His research interests include signal processing, machine learning, tensor analysis, multimodal analysis and network theory. His work has been used for several applications including hyperspectral super-resolution, MRI scan acceleration and multi-view embedding.