Attention in Early Vision: Some Psychophysical Insights

The Spotlight Models of attention that rely upon a bottom-up approach specifically through the dorsal pathways, can be modeled using multi-scale Gaussian pyramids with excitatory-inhibitory feedforward cellular neural networks (CNN) as feature detectors.

  • PDF / 2,289,795 Bytes
  • 18 Pages / 430 x 660 pts Page_size
  • 88 Downloads / 154 Views

DOWNLOAD

REPORT


Abstract. The Spotlight Models of attention that rely upon a bottomup approach specifically through the dorsal pathways, can be modeled using multi-scale Gaussian pyramids with excitatory-inhibitory feedforward cellular neural networks (CNN) as feature detectors. Here we propose a modified disinhibitory zero-feedback CNN model derived out of a linear combination of three Gaussians only, that explains many brightness perception based psychophysical phenomena unexplainable with the old model and in the process predicts three different input cloning templates for global smoothing, global enhancement, as well as controlled smoothing and enhancement of retinal images within the focus of attention. The proposed approach provides new clues, based on the psychophysical stimuli, suggestive of a role of top-down attentional control possibly through the ventral pathways, even at the stage of low-level vision.

1

Introduction

The role of attention in the visual system is to selectively enhance and expedite the processing of a subset of the available stimuli versus the rest. This selective attention in the visual system is frequently implemented in the spatial domain by means of a geometrically well-defined region of the visual field, that is often referred to as the focus of attention (FOA)[1]. Among the various models of attentional processing, there are some that operate on saliency map, which is a topographic representation of the instantaneous saliency of the visual scene, specifying where things and locations of interest are, but not exactly what they are (see the chapter by Bosse et al. in this book). The FOA is computed preattentively and attention then operates on the features inside the FOA for preferential processing. Such features inside FOA are routed to higher processing centers and binding is accomplished by defining that all features inside the FOA belong together. These models are sometimes termed as Spotlight Models (SM). But, in such modeling, a fundamental question remains as to how the attentional spotlight knows where to be directed, i.e. how the saliency map computes the most sensible FOA. The general approach is that, the visual system uses anatomically distinct pathways for encoding spatial information of objects in the environment and the specific features of these objects. While the locations are represented in L. Paletta and E. Rome (Eds.): WAPCV 2007, LNAI 4840, pp. 381–398, 2007. c Springer-Verlag Berlin Heidelberg 2007 

382

K. Ghosh and S.K. Pal

the dorsal or where pathway, the detailed feature processing and object recognition are localized in the ventral or what pathway. If these two pathways function completely independent of one another, then there is hardly any plausible answer to the above problem regarding the process of computation of the saliency map that involves a series of feature detectors working along different feature dimensions and at several spatial scales and then computing the most salient winner. This is because one does not know how clues are obtained as to what the feature di