Band Embedding-Based Network for Spectral Super-Resolution

Spectral super-resolution (SSR) is a means to generate hyperspectral image (HSI) with high-spectral resolution from multispectral image with low-spectral resolution. Most recent works generally exploit the spectral correlation and spatial similarity in HSI to generate the desired HSI, while not considering the spectral similarity between all band images. In this paper, we propose a band embedding-based network (bandEmb-Net), which exploits the high spectral similarity of HSI. We predict the different band images in an HSI with a single network, which contains multiple embeddings of band information to ensure that the output is closer to the desired band image. The proposed bandEmb-Net consists of two parallel branches, including a band embedding block (BEB) and a specific band mapping (SBM). BEB is employed to explore the spectral similarity, making up of several convolution and transformer joint modules. SBM is designed to ensure the unique characterization of each band image in an HSI. Furthermore, the outputs of the BEB and SBM modules are added and followed by a convolutional layer to obtain the final band image. Experimental results on two datasets demonstrate the superior performance of the proposed bandEmbNet.