High-order Spatial and Spatiotemporal Image Statistics and Visual Processing
Author | : Qin Hu |
Publisher | : |
Total Pages | : 202 |
Release | : 2012 |
ISBN-10 | : OCLC:906801358 |
ISBN-13 | : |
Rating | : 4/5 (58 Downloads) |
Book excerpt: This thesis explores the relationship of image statistics to visual processing by considering two important aspects: motion processing and static images. Detection of movement is one of the most fundamental tasks performed by our visual system. The essence of motion is spatiotemporal correlation, but how these correlations are processed biologically is not yet fully known. To probe the computations underlying motion perception, we created a new class of motion stimuli, "glider motion," characterized by their third- and fourth-order spatiotemporal correlations. The direction of glider motion cannot be detected by current models for the neural computation of spatiotemporal correlations. Nevertheless, glider stimuli reliably produced a directional motion percept in humans. This implies that the current models for the computations underlying motion processing require modification, as we discuss. To determine where the calculation of glider motion takes place in the visual system, we recorded from the visual cortex of anesthetized macaque monkeys. The results show that a fraction of V1 and V2 neurons were directionally biased for glider motion, and most of these were also directionally biased for standard motion. We also found something puzzling: for individual neurons, the directional bias for three-element glider motion was opposite to the expectation from the psychophysical results. The second investigation focused on the statistics of natural images. Understanding the design of the visual system and how it extracts information requires knowledge of these statistics. As is well-known, second-order statistics of natural images are characterized by a power spectrum of 1/f 2 . In contrast, higher-order statistics are difficult to study because of their high dimensionality. Two-dimensional Hermite (TDH) functions have mathematical properties that recommend them for analysis of image statistics. We applied TDH functions as filters to a set of natural images, and then calculated the statistics of the filter outputs. The analysis yielded a compact and comprehensive description of high-order statistics of natural images, and this description reflects both their phase and amplitude structure. Finally, we discuss how these findings are related to their "causes" in the real world, and how the visual system can take advantage of them.