Saliency Detection for Hyperspectral Images via Sparse-Non Negative-Matrix-Factorization and novel Distance Measures

Saliency detection is a very active area in computer vision. When hyperspectral images are analyzed, a big amount of data need to be processed. Hence, dimensionality reduction techniques are used to highlight salient pixels allowing us to neglect redundant features. We propose a bottom-up approach based on two main ingredients: sparse non negative matrix factorization (SNMF) and spatial and spectral distances between the input image and the reconstructed one. In particular, we use both well known and novel distance functions. The method is validated on both hyperspectral and multispectral images.