Multilevel Interaction Embedding for Hyperspectral Image Super-Resolution

Single hyperspectral images super-resolution (SHSR) gains a significant achievement with the rapid development of deep learning networks. Hyperspectral images possess a large number of narrow and continuous spectral channels. Traditional methods to process such large number of channels directly often causes the huge parameter and computation of super resolution networks. Group-based methods are proposed to address the dilemma by grouping the hyperspectral image along the spectral dimension into subgroups and super-resolving the image of subgroups. These methods can greatly reduce the parameters and maintain acceptable performance, but existing group-based methods often ignore or execute simple interaction of the subgroups, leading to the weak SHSR performance. In this paper, we propose a multilevel interaction embedding module (MIEM) adapted to the group-based method. MIEM introduces the shallow and deep images feature of the subgroup to the next subgroup to assist super-resolution. The module makes full use of complementarity information among the neighbouring spectral bands to improve SHSR, which can be flexibly applied in group-based methods. The extensive experiments have demonstrated the effectiveness of MIEM.