Edge-finding in human vision: a multi-stage model based on the perceived structure of plaids

Abstract Visual experiments on the perceived structure of plaid patterns are reported and reviewed. Plaids composed of two or three sinusoidal grating components were viewed both with and without the influence of masking or adapting patterns. The results lead to the outline of a multi-stage model of the processes by which human vision derives edge-feature descriptions from the retinal image via local spatial filters. The main proposed stages are: (i) local, oriented filtering; (ii) local combination of filter outputs across orientation (or spatial frequency) to form (iii) filtered image ‘patches’; (iv) coherent combination of patches across space to form a ‘neural image’; (v) zero-crossing analysis on the neural image to localize edge features in (vi) a feature map; (vii) application of coding rules to describe feature attributes at zero-crossing locations. Evidence for the filter combination and zero-crossing stages [(ii) and (v)] comes from experiments on the perceived structure of plaids, where the perceived location and orientation of edges was predicted by zero-crossings in the output of a circular filter, and not by the Fourier component orientations. By contrast, results on the tilt aftereffect for plaids could be understood entirely as an interaction between Fourier components, but with consequential changes in the perception of ZCs that required a consideration of ‘patchwise’ analysis and the patch combination stage [(iii) and (iv)]. Theory and evidence for coding rules (vii) that describe the blur, contrast and orientation of local edge features are discussed.