Image-type dependent eigen-regions-of-interest define conspicuity operators for predicting human scanpath fixation

Top-down informativeness elaboration and bottom-up conspicuity processing are intimately interconnected in visual perception. An internal cognitive model of the external world must necessarily control not only our recognition but also the scanpath sequence of eye movement/shift of attention jumps. A self-organizing process based on principal component analysis and scanpath experimental data are used in this paper to define spatial visual conspicuity from the eigen-features of the scanpath sequence image loci. Eight different classes of images are used as both training and testing set. We first demonstrate that cognitive-driven scanpath loci can be discriminated in terms of these bottom-up eigen conspicuity features. We can finally define a conspicuity processing algorithm and measure its ability to predict human scanpaths as evidenced by the positional similarity measure SpSp. Some computer vision applications will also be discussed.