Orientation Selectivity and Contrast Gain Control in Representations of Natural Images (Orientierungsselektivität und Kontrastverstärkungsregelung in Repräsentationen von natürlichen Bildern)

Fabian H. Sinz

Abstract: This thesis explores the role of orientation selectivity and contrast gain control with respect to Barlow's normative redundancy reduction hypothesis in simple models of the early visual system. Our general approach uses the fact that-under the goal of redundancy reduction-early vision models are density models on natural images. We identify and develop new classes of probabilistic models for natural image patches that contain these early vision models. We use those classes to quantitatively explore their parameter space around the early vision models statistically and information theoretically with respect to the influence of filter shapes and contrast transforms on redundancy reduction. We identify an optimal contrast gain control transform and compare it to the standard model of cortical divisive contrast gain control, divisive normalization. We also identify a new estimation method for the true redundancy of natural images. Our main findings are that, in contrast to divisive contrast gain control, orientation selectivity plays a minor role for redundancy reduction in the models investigated, and that the cortical model of divisive contrast normalization is not the optimal redundancy reducing contrast transformation on static image patches. However, we are able to specify a dynamical model of cortical contrast gain control with strong redundancy reduction, through extending the static model by adaptation to temporal correlations between consecutive contrasts caused by fixations under natural viewing conditions.