Modeling spatial integration and contrast invariance in visual pattern discrimination

Human pattern discrimination performance has been reported to be largely independent of stimulus contrast but to depend on stimulus area. The authors propose a model that combines the effects of spatial integration and contrast. The model is based on the computation of similarity between pattern templates in memory and signals to be discriminated using normalized correlation. There are also two sources of additive noise, one before and one after the computation of correlation. The model was compared with human observers in an orientation discrimination task.

Orientation discrimination thresholds of human observers were measured for sinusoidal gratings of various areas, contrasts, and spatial frequencies. A two-interval, forced-choice methods was used. The performance of the model was determined by using computer simulations.

It was found that the effects of contrast and grating area were interrelated. The decrease of orientation thresholds as a function of grating area was considerably larger at low than at high contrast. On the other hand, orientation thresholds decreased clearly as a function of contrast at the smallest grating areas but hardly at all at the largest grating areas. The model accounted well for the experimental findings.

Because the invariance of orientation discrimination with respect to stimulus contrast depended on area, the cause of the invariance appeared to occur after spatial integration. The model explains this so that, with increasing contrast or area, the normalized correlation gradually approached a constant value. The proportion of pretemplate noise became negligible compared to the constant posttemplate noise. Thus, total noise also approached a constant value. Hence, the signal-to-noise ratio and discrimination performance became constant.