Phase space deep neural network with Saliency-based attention for hyperspectral target detection

Abstract

The accurate separation of targets and background is challenging in hyperspectral target detection algorithms, due to the high variability and complex non-linear scattering interactions in spectra acquired by imaging spectrometers. Moreover, the target regions may be contaminated by the background signal in real images, hindering the separation of a specific target in a scene. To address these challenges, a deep neural network is proposed in this work, consisting of three modules. First, to extract features hidden in the spectral signature of pixels, the hyperspectral image is considered as a dynamic system, and its phase space is reconstructed in the spectral feature space. Subsequently, in order to highlight the targets and suppress the background, a saliency map is produced, which shows candidate regions for the targets of interest. The saliency map is then utilized as an attention map for weighting the hyperspectral input within the network. The proposed multi-branch deep neural network processes each dimension of the reconstructed phase space. The resulting Phase Space Deep Neural Network with Saliency-based Attention (PSDNN-SA) outperforms several state-of-the-art detectors both quantitatively and visually in experiments carried out on different real hyperspectral subsets.