Visual statistics using neural networks.

Abstract

This thesis describes the application of statistical techniques to natural images as a means of gaining insight into the operation of low level vision. First, the statistical technique of principal component analysis is applied to a collection of natural images: a match with psychophysical data is found; and a solution to the dynamic range problem proposed. The problem of learning and calibrating psychological and physiological representations of space is t hen investigated. The grey level correlations in natural images are measured and their physical causes investigated. The resulting correlations are related both to psychological distortions of space and to the cortical representation of space in \T in macaque monkey. The interpretation in terms of a system calibrating itself using the correlations in the input signals is shown to produce accurate psychological and physiological predictions. Lastly the problems of creating low level models of the visual input is looked at using a framework originally proposed by Hinton and Sejnowski (1983). The way in which phase coherence of (neuronal) firing in a network can label the probability of an interpretation is demonstrated. A new search technique, inspired by the different time courses of inhibition and excitation in the cortex, is proposed for searching for the most likely visual interpretation. It is concluded that statistical techniques can provide insight into the operation of low level vision.