There is an inherent degree of uncertainty involved in making perceptual decisions. This uncertainty may come from ambiguous causes of sensory input, or from random variation in the physical or biological processes involved in transmitting perceptual information. Human observers are remarkably good at making valid and reliable visual decisions despite random variation in the image, compared to capabilities of most artificial vision systems. This project addresses how human observers process ambiguous or "noisy" visual information. Performance can be calculated for a theoretically "optimal" or "ideal" observer for a visual task that has inherent uncertainty. The optimal performance assumes ideal performance from the optics and detector elements of the system, and the calculations are based on a branch of probability theory called statistical decision theory. By measuring the visual performance of a human and comparing it to the ideal benchmark, we can measure human absolute visual efficiency. Patterns which are detected nearly optimally may reflect specialized neural mechanisms for a particular task. Here the role of spatial arrangement, color, texture, and motion will be studied to find out what types of images are detected most efficiently. The limits to optimal coding of more natural images will be quantified by measuring the predictability of various elements in pictures. This study is an unusual combination of experimental work with very sophisticated mathematical modelling, and will help integrate visual perception theory with both machine vision and human performance. It should have substantial impact on vision research.
