Hyperspectral Image Super-Resolution via Self-projected Smooth Prior

Abstract

High spectral correlations and non-local self-similarities, as two intrinsic characteristics underlying hyperspectral image (HSI), have been widely used in HSI super-resolution. However, existing methods mostly utilize the two intrinsic characteristics separately, which still inadequately exploit spatial and spectral information. To address this issue, in this study, a novel self-projected smooth prior (SPSP) is proposed for the task of HSI super-resolution. SPSP describes that two full-band patches (FBPs) are close to each other and then the corresponding subspace coefficients are also close to each other, namely smooth dependences of clustered FBPs within each group of HSI. Suppose that each group of FBPs extracted from HSI lies in smooth subspace, all FBPs within each group can be regarded as the nodes on an undirected graph, then the underlying smooth subspace structures within each group of HSI are implicitly depicted by capturing the linearly pair-wise correlation between those nodes. Utilizing each group of clustered FBPs as projection basis matrix can adaptively and effectively learn the smooth subspace structures. Besides, different from existing methods exploiting non-local self-similarities with multispectral image, to our knowledge, this work represents the first effort to exploit the non-local self-similarities on its spectral intrinsic dimension of desired HSI. In this way, spectral correlations and non-local self-similarities of HSI are incorporated into a unified paradigm to exploit spectral and spatial information simultaneously. As thus, the well learned SPSP is incorporated into the objective function solved by the alternating direction method of multipliers (ADMM). Experimental results on synthetic and real hyperspectral data demonstrate the superiority of the proposed method.