Integrating Inhomogeneous Processing and Proto-object Formation in a Computational Model of Visual Attention

Abstract

We implement a novel computational framework for attention that includes recent experimentally derived assumptions on attention which are not covered by standard computational models. To this end, we combine inhomogeneous visual processing, proto-object formation, and parts of TVA (Theory of Visual Attention [2]), a well established computational theory in experimental psychology, which explains a large range of human and monkey data on attention. The first steps of processing employ inhomogeneous processing for the basic visual feature maps. Next, we compute so-called proto-objects by means of blob detection based on these inhomogeneous maps. Our model therefore displays the well known ”global-effect” of eye movement control, that is, saccade target landing objects tend to fuse with increasing eccentricity from the center of view. The proto-objects also allow for a straightforward application of TVA and its mechanism to model task-driven selectivity. The final stage of our model consists of an attentional priority map which assigns priority to the proto-objects according to the computations of TVA. This step allows to restrict sophisticated filter computation to the proto-object regions and thereby renders our model computationally efficient by avoiding a complete standard pixel-wise priority computation of bottom-up saliency models.