Hyperspectral Anomaly Detection Based on Multicomplementary Prior-Guided Tensor Decomposition

Recently, tensor decomposition technology has been widely introduced into hyperspectral anomaly detection (HAD), aiming to exploit the data characteristics of the background tensor fully. Despite this, the existing tensor-decomposition-based methods do not fully dig into the various internal priors of the background (i.e., the low-rankness and piecewise-smoothness of the background in three dimensions, the nonlocal self-similarity) as well as the external prior (i.e., the deep prior learned from external image datasets). In this article, we proposed a multicomplementary prior-guided tensor decomposition (MCPGTD) method for HAD. Specifically, a learnable dictionary with double constraint is introduced to exploit the low-rankness and piecewise-smoothness of the background in spectral dimension; the low-rankness, nonlocal self-similarity, and piecewise-smoothness of the background in spatial dimension are exploited by designing double complementary constraint on the coefficient tensor. Besides, a plug-and-play (PnP) trained network is plugged into this tensor to introduce the external deep prior. After all complementary regularizations have been considered, an efficient alternating direction method of multipliers (ADMMs)-based approach is introduced to solve the variables included in the proposed model. We finally prove the effect detection performance of the proposed model on several real hyperspectral datasets.