Novel Reconstruction Errors for Saliency Detection in Hyperspectral Images

When hyperspectral images are analyzed, a big amount of data, representing the reflectance at hundreds of wavelengths, needs to be processed. Hence, dimensionality reduction techniques are used to discard unnecessary information. In order to detect the so called “saliency”, i.e., the relevant pixels, we propose a bottom-up approach based on three main ingredients: sparse non negative matrix factorization (SNMF), spatial and spectral functions to measure the reconstruction error between the input image and the reconstructed one and a final clustering technique. We introduce novel error functions and show some useful mathematical properties. The method is validated on hyperspectral images and compared with state-of-the-art different approaches.