Local Energy Saliency for Bottom-Up Visual Attention

Visual attention is a key visual process of selection of information, with a wide variety of applications in vision systems and image processing. In this paper we hold and test a novel model based on the Koch and Ullman architecture of saliency-based attention. We propose a novel feature from maximum local energy, already used in edge and feature extraction, but in ways which aren't suitable in attention applications. The envelope of a bank of log Gabor filters (resembling the receptive fields of complex cells from visual cortex) is taken as a local energy measure. In our approach to the feature integration problem, across the scales of each orientation, we compute initial conspicuity maps employing the T2 Hotelling statistic, as a multivariate measure of variance, and taking by this way conspicuous points with maximum local energy. Normalizing and integrating these maps, gathering the highest values of variance, we obtain a unique final saliency measure. With this model we achieve improved results accounting for orientation pop out and equivalent performance in a visual search task within cluttered natural scenes, both in comparison with analogue experiments previously published with a state of art model

keywords: Visual Attention, Bottom-up, Saliency, Feature Integration, PCA, Local Energy