Séminaire de Yinjian Wang

Speaker: Yinjian Wang, PhD student at the Beijing Institute of Technology
Title: Degradation-Modeling-Induced Tensor Inverse Problems
Link: https://rendez-vous.renater.fr/Seminaire_Yinjian_242bff-c5b9a5-0cff34

Abstract: Tensor inverse problems are central to multidimensional scientific data reconstruction. This seminar discusses how explicit degradation modeling can reshape the formulation, theory, and algorithms of tensor-based reconstruction. In hyperspectral image super-resolution, conventional tensor models usually assume separable spatial degradation, which may fail under realistic sensor effects such as anisotropic blurring. A generalized tensor formulation based on Kronecker decomposition is introduced to handle arbitrary spatial-degradation matrices and to establish recoverability conditions for exact reconstruction.

For foreground-background separation, classical robust principal component analysis assumes additive sparse corruption, whereas real foreground objects often replace or occlude the background. This motivates a support-aware completion formulation, where the resulting NP-hard support estimation problem is addressed through Bayesian sparse tensor factorization and variational Bayesian inference, leading to probabilistic support estimation and threshold-free hard classification.

Together, these studies show that degradation modeling is not merely an observation detail, but a key factor shaping tensor inverse problems.