Abstract
Whereas receptive fields of the primary visual cortex have been extensively characterized, the integration mechanisms constructing a global percept from these local computations are largely unknown. Additionally, it has been suggested that the visual system is tuned to the statistics of the external environment. We tested what statistical properties of visual stimuli may underlie integration processes. Specifically, we examined the sensitivity of human vision to order-disorder transition in visual textures. We employed a set of textures generated from different homogeneous Markov Random Fields (MRF). Changing a one-dimensional parameter (analogous to the thermodynamic temperature in Boltzmann distribution), the generated textures vary from random (independent identically distributed amplitudes, IID) to ordered. An order parameter, used in statistical physics to identify order-disorder transition was estimated for each MRF. We measured psychometric function (performance vs. the MRF parameter) in a 4AFC task. In each display one texture was controlled by the MRF parameter while the other three were random textures. Observers (n=7) were requested to report in which quadrant of display the most ordered texture appeared. Human performance followed the order parameter curve for MRF where many visually different images corresponded to the same ordered state. Notably, for MRFs with a single image corresponding to an order state (prototype image), such as vertical gratings or checkerboards, a lower level of order was required to identify the texture as ordered. All observers were substantially outperformed by an ideal observer based on bootstrapped likelihood ratios between sufficient statistics. Our results suggest that integration mechanisms in the visual system effectively compute an order parameter representing qualitative changes in images, such as order-disorder. A visually prototyped image enabled enhanced order sensitivity. Moreover, the human visual system lacks the flexibility required to form an efficient representation (sufficient statistics) for the given task.
Meeting abstract presented at VSS 2014