Early stages in spatial vision

Despite the ease with which we perceive, it is not clear how the distribution of light across the visual field is analyzed by eye and brain to give rise to our meaningful percepts. One approach to study visual perception is to focus on the very first stages of information processing. Without precise knowledge of the computational processes that take place in early vision, it is difficult, if not impossible, to understand the more complex later stages in detail. Much evidence has accumulated that suggests that the early visual system is comprised of spatial frequency-selective channels. It is generally assumed that in contrast detection circumstances, these channels behave linearly. This assumption has been challenged by one study that reported contrast detection facilitation in the presence of weak levels of luminance noise (Blackwell, 1998). Improved information transmission in the presence of certain levels of externally added noise, i.e. stochastic resonance, is the signature of nonlinear information processing systems. Contrast detection facilitation in the presence of weak levels of 2-D, white luminance noise is explored in Part I. Comparison of contrast thresholds measured in two different tasks, namely signal detection in Experiment la and orientation discrimination in Experiment lb, but in otherwise identical circumstances, demonstrates that the origin of this facilitation effect is not to be found in a higher level task strategy. Minimization of spatial and temporal uncertainty about signal occurrence and the provision of feedback in Experiment 2 suggest that the visibility of low contrast signals is truly enhanced in the presence of weak noise levels. It is demonstrated that these data are consistent with a model of our visual system wherein a linear filter is followed by a nonlinear post-filter stage. The implication of this finding for critical band masking, a technique to infer filter shape from detection in noise data is explored in Part II. In Experiment 3, by making use of low-pass filtered noise, i.e. noise from which all high frequency components have been removed, it is shown that the noise must not be white for detection facilitation to occur. Activation of the signal processing channel by the noise is sufficient. Experiment 4 demonstrates empirically that the power spectral density of the noise does matter in the critical band masking paradigm: Different noise levels lead to different results. A novel way to infer filter shape from these data is suggested. Sinusoidal contrast discrimination in the presence of luminance noise is explored in Part III. Experiment 5 demonstrates that the well-known dipper-shaped threshold-vs.-contrast function (TvC function) changes somewhat in the presence of weak and strong noise levels. To be able to consider the effects of noise on contrast discrimination, irrespective of performance level, a computational contrast perception model was fitted to the data. Analysis of model deviances reveals that addition of weak noise diminishes the facilitating effect of a low contrast masker. Addition of strong noise causes a small rightward shift of the TvC function on double logarithmic coordinates. Whatever mechanism underlies contrast discrimination, future contrast perception models should aim to be able to explain these findings. Recent work has questioned the validity of the current set of models of human spatial vision (Henning & Wichmann, 2007). The framework of a possible new model which might overcome some of the modelling problems is presented in Part IV. The central idea is to combine V1 neuron-like spatial frequency filters with a population output decision rule. Via simulations it is shown that this results in a model which reconciles constant tuning of the building blocks, as has been postulated often for cortical cells, with dynamic tuning properties of psychophysical channels.status: publishe