Global Precedence, Spatial Frequency Channels, and the Statistics of Natural Images

摘要

A great deal of evidence suggests that early in processing, retinal images are filtered by parallel, spatial frequency selective channels. We attempt to incorporate this view of early vision with the principle of global precedence, which holds that Gestalt-like processes sensitive to global image configurations tend to dominate local feature processing in human pattern perception. Global precedence is inferred from the pattern of reaction times observed when visual patterns contain multiple cues at different levels of spatial scale. Specifically, it is frequently observed that global processing times are largely unaffected by conflicting local cues, but local processing times are substantially lengthened by conflicting global cues. The asymmetry of these effects suggests the dominant role of global configurations. Since global spatial information is effectively represented by low spatial frequencies, global precedence potentially implies a low frequency dominance. The thesis is that low spatial frequencies tend to be available before information carried by higher frequency bands, producing a coarse-to-fine temporal order in visual spatial perception. It is suggested that a variety of factors contribute to the “prior entry” of low frequency information, including the high contrast gain of the magnocellular pathway, the amplitude spectra typical of natural images, and inhibitory interactions between the parallel frequency-tuned channels. Evidence suggesting a close relationship between global precedence and spatial frequency channels is provided by observations that the essential features of the global precedence effect are obtained using patterns consisting of low and high frequency sinusoids. The hypothesis that these asymmetric interference effects are due to interactions between parallel spatial channels is supported by an analysis of reaction times (RTs), which shows that RTs to redundant low and high frequency cues produce less facilitation than p- edictions that assume the channels are independent. In view of previous work showing that global precedence depends upon the low frequency content of the stimuli, we suggest that low spatial frequencies represent the sine qua non for the dominance of configurational cues in human pattern perception, and that this configurational dominance reflects the microgenesis of visual pattern perception. This general view of the temporal dynamics of visual pattern recognition is discussed, is considered from an evolutionary perspective, and is related to certain statistical regularities in natural scenes. Potential adaptive advantages of an interactive parallel architecture that confers an initial processing advantage to low resolution information are explored.