Hyperspectral Image Superresolution via Deep Structure and Texture Interfusion

Limited by the imagery sensors, hyperspectral image (HSI) is characterized by its high spectral resolution, but low spatial resolution. HSI superresolution (SR) aims at enhancing the spatial resolution through the postprocessing techniques. Inspired by the important observation that the structure and the texture exhibit different sensitivities to the spatial resolution, we propose an HSI SR method via a residual structure-texture dense network. Specifically, there are three modules that are designed in the SR process. First, the input HSI is decomposed into the texture cube and the structural cube via relative total variation. Second, the mapping between the texture-structure cube of the input HSI and that of the desired HSI is learnt via two parallel subnetworks. Finally, the deep learnt texture cube is interfused with the deep learnt structural cube to achieve the reconstructed HSI with high spatial resolution via a upsample module. Contrary to the existed methods that mainly focus on the entire scene, our approach opens a new way for superresolving the HSI from a structure and texture recovery aspect. Experimental results and data analysis on both indoor scenes and outdoor scenarios have demonstrated the effectiveness of the proposed method.