Abstract
Visual saliency is the perceptual quality that makes some items in visual scenes stand out from their immediate contexts. Visual saliency plays important roles in natural vision in that saliency can direct eye movement, deploy attention, and facilitate object detection and scene understanding. Natural visual scenes consist of objects of various physical properties that are arranged in three dimensional space in a variety of ways. When projected onto the retina, visual scenes entail highly structured statistics, occurring over the full range natural variation in the world. Thus, a given visual feature could appear in many different ways and in a variety of contexts in natural scenes. Dealing effectively with these enormous variations in visual feature and their contexts is a paramount requirement for routinely successful behaviors. Thus, for visual saliency to have any biological utility for natural vision, it has to tie to the statistics of natural variations of visual features and the statistics of co-occurrences of natural contexts. Therefore, we propose to explore and test a novel, broad hypothesis that visual saliency is based on efficient neural representations of the probability distributions (PDs) of visual variables in specific contexts in natural scenes, referred to as context-mediated PDs in natural scenes. We first develop efficient representations of context-mediated PDs of a range of basic visual variables in natural scenes. We derive these PDs from the Netherland database of natural scenes and the McGill dataset of natural color images using independent component analysis. We then derive a measure of visual saliency based on context-mediated PDs in natural scenes. Experimental results show that visual saliency derived in this way predicts a wide range of perceptual observations related to texture perception, pop-out, saliency-based attention, and visual search in natural scenes.