The spatial area of the normalization pool in early visual processing

Precisely measuring the properties of low-level visual processing is of theoretical interest as such properties often generalize to other processing in the brain and have immediate applications for image and video processing. The initial stages of human visual processing split the signal into spatial frequency and orientation-tuned channels, which interact nonlinearly, are typically modeled as divisive normalization. While we know in detail which spatial frequency and orientation channels interact, we know much less about how far the interactions reach across space. In particular, we lack psychophysical data to constrain the spatial area of the normalization pool. To provide such constraints, we set up forced choice masking experiments varying the size of the mask while we keep the spatial frequency and orientation of the mask the same as the targets. The target grating was always a small (0.5 cycles standard deviation) circular Gabor centrally superimposed over 15% and 40% contrast masks. We tested eight different mask sizes for 1, 2, and 4 cycles per degree vertical stimuli in the fovea and at 5° and 10° eccentricities. In the periphery at 5° and 10° eccentricities, contrast thresholds increase with increasing mask size and then saturate as expected, which allows us to estimate the spatial extent of the normalization pool by fitting a divisive normalization model. The fitted Gaussian normalization pool has a standard deviation of 0.65° for 1 cycle per degree gratings and scaled inversely proportional to the spatial frequency of the gratings. Neither the eccentricity nor the mask contrast influenced the size of the normalization pool. In the fovea, however, the size of the mask had no influence on thresholds, even though the mask reliably elevated thresholds. Thus, we are left with an accurate estimate of normalization pool sizes in the periphery, along with the unresolved question of why masking in the fovea shows a different behavior.