A deep image prior-based interpretable network for hyperspectral image fusion.

Hyperspectral image fusion aims to generate a high-resolution hyperspectral image by utilizing a low-resolution hyperspectral image and a high-resolution multispectral image. Inspired by model-based image fusion method, we propose a new image fusion network which use deep prior information as spatial guide information. Our fusion network is mainly obtained by the following steps: First, we propose a fusion model which use spatial information as regularization term. Second, we use half splitting quadratic method to solve the improved fusion model. Third, we interpret the model with a network structure. In the network, we use encoding-decoding structure to construct the spatial regularization term. Finally, in order to ensure both spectral and spatial information contained in each iteration can be fully utilized, an aggregation module is added at the end of the network. Experimental results on three simulation data sets demonstrate that the proposed method performs well in solving the fusion problem. [ABSTRACT FROM AUTHOR]