Models of Bottom-Up Attention and Saliency

Visually conspicuous, or so-called salient, stimuli often have the capability of attracting focal visual attention towards their locations. Several computational architectures subserving this bottom-up, stimulus-driven, spatiotemporal deployment of attention are reviewed in this article. The resulting computational models have applications not only to the prediction of visual search psychophysics, but also, in the domain of machine vision, to the rapid selection of regions of interest in complex, cluttered visual environments. We describe an unusal such application, to the objective evaluation of advertising designs. One of the most important functions of selective visual attention is to rapidly direct our gaze towards objects of interest in our visual environment. From an evolutionary standpoint, this rapid orienting capability is critical in allowing living systems to quickly become aware of possible preys, mates or predators in their cluttered visual world. It has become clear that attention guides where to look next based on both bottom-up (image-based) and top-down (task-dependent) cues (James, 1890/1981). As such, attention implements an information processing bottleneck, only allowing a small part of the incoming sensory information to reach short-term memory and visual awareness [Linking Attention to Learning, Expectation, Competition and Consciousness]. That is, instead of attempting to fully process the massive sensory input in parallel, nature has devised a serial strategy to achieve near real-time performance despite limited computational capacity: Attention allows us to break down the problem of scene understanding into rapid series of computationally less demanding, localized visual analysis problems. Developing computational models of how attention is deployed onto complex visual scenes has been a longstanding challenge for fundamental neuroscience, with additional motivation provided by numerous potential applications in artificial vision, for tasks including surveillance, automatic target detection, navigational aids and robotics control. Here we focus on biologically-plausible computational modeling of bottom-up guidance of attention towards salient image locations, while [Attention and scene understanding] casts these models within broader frameworks that combine bottom-up and top-down attention control signals. 1 Preattentive Features and Saliency Map Development of computational models of attention started with the Feature Integration Theory of Treisman & Gelade (1980), which proposed that only simple visual features are computed in a massively parallel ∗University of Southern California, Hedco Neuroscience Building HNB-30A, Los Angeles, CA 90089-2520