Figure 6.

A diagrammatic representation of the oriented difference-of-Gaussian (ODOG) model. (a) A gray level representation of an ODOG filter. The oriented filters of the ODOG model are produced by setting the ratio of DOG center/surround space constants to 1:2 in one orientation and to 1:1 in the orthogonal orientation. (b) The ODOG model is implemented in 6 orientations (0, 30, 60, 90 −30 and −60 degrees relative to vertical). Each orientation is represented by seven volume-balanced (i.e., integrate to 0) filters that possess center frequencies arranged at octave intervals (from 0.1–6.5 c/d). The seven filters (b) within each orientation are summed after weighting across frequency using a power function with a slope of 0.1 (c). This slope is consistent with the shallow low-frequency fall-off of the suprathreshold contrast sensitivity function (Georgeson and Sullivan, 1975). The resulting six multiscale spatial filters, one per orientation, are convolved with the stimulus of interest (d–e). The filter outputs (f) are normalized across orientation by dividing each by its space-averaged root-mean-square contrast, as computed across the entire convolution output (g). The six normalized outputs are summed to produce the final ODOG model output (h).