Nonlocal Tensor Wheel Decomposition for Hyperspectral Image Super-Resolution

Abstract

Fusing paired multispectral image (MSI) with hyperspectral image (HSI) has emerged as a prevalent scheme for HSI super-resolution (HSR). We propose a nonlocal tensor wheel decomposition (NLTW) approach for HSR. By introducing TW decomposition for nonlocal group representation, the proposed NLTW model effectively explores the nonlocal self-similarity prior. Compared with existing nonlocal tensor decompositions for HSR, our model leverages higher-order representations and establishes connections among non-adjacent factors, exhibiting more expressive characterization capability. Moreover, we develop an efficient algorithm based on the alternating direction multipliers method (ADMM) and proximal alternating minimization (PAM), with Bayesian optimization strategy for automated hyperparameter tuning. Experiments on three datasets demonstrate the superiority of our model over state-of-the-art methods