Abstract
The perceived brightness of a surface depends on the brightness of neighboring surfaces. Here, we describe a new low-level computational model of this effect. We extend the ODOG model (Blakeslee & McCourt, 1999), which combines two simple mechanisms. First, the input is filtered by multiscale oriented difference of Gaussian filters. Second, global response normalization equalizes the amount of energy at each orientation across the entire visual field.
In this work we extended the ODOG model with a more neurally plausible normalization step. The normalization step in the ODOG model is necessary to account for a family of illusions known as White's effect, which are often characterized by a highly non-uniform distribution of energy at different orientations. ODOG fails on variations of White's effect that have equal energy across orientations when integrated over the entire image, suggesting a more localized normalization scheme is necessary. A local mechanism also has the advantage of being more plausible for implementation in early visual areas, such as V1, because it only requires short-distance connections between neurons.
In our new model, Frequency-specific Locally-normalized ODOG (FLODOG), energy normalization is computed locally, both in terms of spatial location and spatial frequency. We filter the image into 6 different orientations and 7 scales. Each filter response is normalized by a weighted sum that includes itself and also filter responses for nearby spatial frequencies of the same orientation. This normalization occurs within a local window, the size of which scales with the spatial extent of the filter being normalized.
The FLODOG model successfully accounts for most of the illusions for which ODOG makes correct predictions. In addition, it correctly predicts many variants of White's illusion that ODOG cannot.