A novel model of bottom-up visual attention using local energy

In this paper we propose a novel model for the implementation of the Koch & Ullman architecture of bottom-up visual attention. We use two features to measure saliency: local energy and color. Local energy is extracted as the envelope of a bank of complex-valued log Gabor filters (resembling the receptive fields of complex cells from visual cortex), giving regions of maximum phase congruency (PC) in the image, which have proven perceptually relevant. Previous PC measures have been employed in edge extraction but they aren't suitable in attention applications. In our approach, across the scales of each orientation, we compute the T 2 Hotelling statistic, as a multivariate measure of variance, that is of maximum PC, and we take it as a conspicuity map. We forward normalize and integrate these maps with color maps to form a unique final saliency measure, taking the highest values of variance at each point. With this model we reach improved or similar performance in comparison to a state of art model, when we reproduce the same test experiments.

keywords: attention, bottom-up, saliency, feature integration, principal component analysis