Hyperspectral image super-resolution via multi-domain feature learning

Hyperspectral image super-resolution (SR) methods have achieved great success due to deep neural networks. Despite this, these methods hardly utilize more 2D convolutions to explore more spatial features when the spectral information can be extracted. Besides, they do not make full use of multi-domain features to realize information complementation. To address these issues, we propose a multi-domain feature learning network using 2D/3D convolution for hyperspectral image SR. A multi-domain feature learning strategy is proposed to explore the spatial and spectral knowledge by sharing spatial information. To better fuse those feature from different domains, the multi-domain features fusion module is introduced to learn more effective information, so as to further realize information complementation. Moreover, to recover the more edge details, we design the edge generation mechanism to explicitly enable the network provide priori edge. Extensive experiments on two benchmark datasets show that the proposed approach produces the state-of-the-art results over the existing works in terms of spatial reconstruction and spectral fidelity.