An Improved Camouflage Target Detection Using Hyperspectral Image Based on Block-Diagonal and Low-Rank Representation

Abstract

Accurate camouflage target distinction is often resorted to hyperspectral spectral imaging technique as for the rich spectral information contained in hyperspectral images. In this paper, a novel block-diagonal representation based camouflage target detection method is proposed for hyperspectral imagery. To better represent the multi-mode cluster background, an hyperspectral image is first clustered into different background clusters according to their spectral features. Then, an orthogonal background dictionary is learned for each cluster via a principle component analysis (PCA) learning scheme. The background and camouflage target often show different structures when projected onto those dictionaries. The former exhibits block-diagonal structure while the latter shows sparse structure. Inspired by this fact, we cast the block-diagonal structure into a low-rank representation model. With proper optimization of such model, the sparse camouflage targets can be accurately separated from the block-diagonal background. Experimental results on the real-world camouflage target datasets demonstrate that the proposed method outperforms the state-in-art hyperspectral camouflage target detection methods.