Early Stages in Spatial Vision - Psychophysical Data and Computational Models of Contrast Perception

All current models of spatial vision postulate an initial linear spatial frequency- and orientation-selective filtering stage. Because sinusoids are the eigenfunctions of any linear system, sinusoidal gratings have b ecome perhaps the single most popular stimuli to investigate early visio n. Often, visual noise is added to these stimuli in behavioural experime nts designed to characterize certain aspects of the visual mechanisms me diating pattern vision. External noise methods are popular because noise is a random stimulus and thus able to yield unbiased results , yet i ts characteristics are highly controllable. However, effects of visual n oise on the processing of sinusoidal gratings are not well understood. A t the neurophysiological level, there have been few attempts to systemat ically explore contrast processing in the presence of visual noise. At t he psychophysical level, assumptions regarding noise effects vary from e xcitation to inhibition. The central goal of this thesis is to advance o ur understanding of the visual processing of sinusoidal gratings embedde d in visual noise. To this end, behavioural measurements of sinusoidal c ontrast detection and discrimination in noise are conducted and computat ional models of contrast perception in spectrally flat and filtered nois e are developed. The most prominent computational models developed to account for contras t detection in noise all predict linear threshold-vs.-noise-contrast functions. Here, it is argued that this prediction is inconsistent with the standard model of the early visual system as derived from sinusoidal contrast discrimination experiments and neurophysiological findings. In the standard model, the linear filtering stage is followed by a (rectif ying) nonlinearity and internal noise. Such linear-non-linear cascade mo dels predict an improvement in observers contrast detection performance when low, sub-threshold levels of external noise are added (i.e., stoch astic resonance). Carefully designed behavioural measurements of contras t detection in noise were conducted in two experiments. Detection thresh olds were found to reach a minimum for very low noise levels. It is demo nstrated that an elaborated version of the divisive contrast-gain contro l model fits these data reasonably well. In the gain-control model, a single mechanism underlies stochastic reson ance and the pedestal effect the pedestal effect is the improvement in the detectability of a sinusoidal grating in the presence of another lo w-contrast grating of the same orientation, spatial frequency and phase. An alternative hypothesis may be to argue that both effects characteriz e two independent mechanisms. To address this issue, sinusoidal contrast discrimination in weak and modest levels of noise was measured in a thi rd experiment. Making use of a full quantitative description of the data combined with comprehensive model selection assessments, the pedestal e ffect is shown to be more reduced in the presence of weak noise than in moderate noise. This reduction rules out independent, additive sources o f performance improvement and supports the parsimonious explanation that a single mechanism underlies the pedestal effect and stochastic resonan ce in contrast perception. A simulation study demonstrates that the elab orated gain-control model can produce similar results. Recent evidence has demonstrated that the pedestal effect in spatial vis ion is differently modified by spectrally flat and notch-filtered noise: the pedestal effect is reduced in flat noise, but virtually disappears in the presence of notched noise (Henning & Wichmann, 2007). As all othe r current models of spatial vision, the standard gain-control model fail s to account for these results. A new population-code model for human co ntrast discrimination is proposed. It is demonstrated that combining the outputs of many units whose contrast response functions resemble those of the cortical cells believed to underlie human pattern vision by simpl e response-based pooling, produces contrast-discrimination data consiste nt with psychophysical observations: the pedestal effect is present with out noise, reduced in broadband noise, but almost disappears in notched noise. These findings follow naturally from the normalization model of s imple cells in primary visual cortex followed by response-based pooling, and suggest that in processing even low-contrast sinusoidal gratings, t he visual system may combine information across neurons tuned to differe nt spatial frequencies and orientations.status: publishe