부산대학교 도서관

Model-driven tensor robust principal component analysis (TRPCA) has been widely applied to feature extraction of hyperspectral images (HSIs) and successfully protected two-dimensional spectral contextual information. Nevertheless, the current TRPCA-based feature extraction methods still destroy the underlying spectral and spatial–spectral joint contextual features. Moreover, these global iterative algorithms commonly ignore the heterogeneity of different real-world regions, increase the calculation burden, and improve practice operating time. To solve these issues, an efficient reconstructive homogeneous block-based local TRPCA is proposed for low-rank feature extraction, composed of a homogeneous block rebuilder and a local TRPCA low-rank feature extractor. The proposed local TRPCA is a novel data-model-driven algorithm depending on the data regulation. It remains the primary spatial and spectral contextual information and extracts the underlying homogeneity and heterogeneity characteristics of spatial, spectral, and spatial–spectral joint variables, which provides more essential features for further research than other model-driven TRPCA models. Furthermore, our local TRPCA feature extractor is an elaborate divide-and-rule model that executes on each homogeneous data block to extract low-rank features adaptively, remarkably decreasing computing cost and time. Experimental results on six hyperspectral datasets demonstrate that the proposed local TRPCA is more adaptive to HSIs and outperforms other state-of-the-art TRPCA-based feature extraction algorithms.