Modelling sparse saliency maps on manifolds: Numerical results and applications

Abstract

Saliency detection is an image processing task which aims at automatically estimating visually salient object regions in a digital image mimicking human visual attention and eyes fixation. A number of different computational approaches for visual saliency estimation has recently appeared in Computer and Artificial Vision. Relevant and new applications can be found everywhere varying from automatic image segmentation and understanding, localization and quantification for biomedical and aerial images to fast video tracking and surveillance. In this contribution, we present a new variational model on finite dimensional manifolds generated by some characteristic features of the data. A Primal-Dual method is implemented for the numerical resolution showing promising preliminary results.