Preattentive discrimination of relative phase modelled by interacting Gabor or by difference-of-Gaussian filters

A computer simulation of the human preattentive visual pathway used Gabor and difference-of-Gaussian (DOG) filters to model two-dimensional relative phase discrimination. There is a hierarchy of hexagonally packed levels: (i) an image layer; (ii) an intermediate level of on- and off-centre DOG filters; (iii) a partial set of broadband oriented Gabor-like filters with high-level DOG filters in parallel. Connections between layers use half-wave rectification and a compressive nonlinearity. Local feedback interactions between oriented Gabor filters, together with spatial averaging, allow the model to discriminate both two-dimensional relative phase and orientation differences. There were two separate simulations for the Gabor filters, one with even-symmetric filters and another with odd-symmetric. Textures were formed by superposing three high contrast sine-wave gratings with successive rotations of 60 degrees. Relative phase and global orientation were the varied parameters. Psychophysical rating data for peripheral viewing of texture pairs resemble the results from the even-symmetric simulation, showing discrimination of relative phase and orientation. In contrast, the DOG filters in the model simulate only the relative phase aspects of the data.