Tensor adaptive reconstruction cascaded with spatial-spectral fusion for marine target detection

Abstract

Hyperspectral target detection has a wide range of applications in marine target monitoring. Traditional methods for target detection take less consideration of the inherent structural information of hyperspectral images and make insufficient use of spatial information. These algorithms may experience degradation in efficacy during complex scenarios. To address these issues, this study introduces a hyperspectral target detection approach based on tensor adaptive reconstruction cascade spatial-spectral fusion, named as TRSSF. First, the position of the pixel that best matches the prior spectrum is obtained. Second, tensor decomposition and reconstruction of the original hyperspectral data are performed. Linear total variation smoothing is used to acquire the principal components in the spatial dimensionality unfolding of data, and correlation regularization robust principal component analysis is employed to derive the spectral dimensionality unfolding’s principal components of data. Finally, the spatial-spectral fusion method is proposed for detecting hyperspectral targets on the reconstructed data. The use of multi-morphological feature fusion can fully utilize the spatial features to complement the spectral detection results and improve the integrity of target detection. The experiments conducted on the publicly available dataset and collected datasets demonstrated the effective detection achieved by the proposed method.