Saliency Detection in Hyperspectral Images Using Autoencoder-Based Data Reconstruction

Saliency detection extracts objects attractive to a human vision system from an image. Although saliency detection methodologies were originally investigated on RGB color images, recent developments in imaging technologies have aroused the interest in saliency detection methodologies for data captured with high spectral resolution using multispectral and hyperspectral imaging (MSI/HSI) sensors. In this paper, we propose a saliency detection methodology that elaborates HSI data reconstructed through an autoencoder architecture. It resorts to (spectral-spatial) distance measures to quantify the salience degree in the data represented through the autoencoder. Finally, it performs a clustering stage in order to separate the salient information from the background. The effectiveness of the proposed methodology is evaluated with benchmark HSI and MSI data.