Abstract
Introduction. Shape perception is a critical precursor to object recognition. Humans reach hyperacuity levels when discriminating shape distortions in the form of radial frequency (RF) patterns from circles, and the shape of the threshold function suggests the use of global cues (Wilkinson, Wilson & Habak, 1998, Vision Research). Here, we developed a neural model of RF perception based on known V4 properties that exhibits many of the characteristics of human RF perception. Model. The first stage of our model uses large-scale non-Fourier V4-like concentric units to estimate the center of ellipsoidal contours by encoding the center of concentric contour segments across orientations. The second stage of our model uses curvature detectors applied directly to contour energy, around the pattern center estimated from the first stage. Curvature responses were highest at points of maximum curvature, thus encoding their number, amplitude, and locations. This population code for shape was analyzed as a Fourier spectrum of curvature responses as a function of RF, where ‘pure’ RF patterns were usually encoded as a single peak amplitude at the corresponding RF. That spectrum was contrasted for test pattern and a comparison pattern (either a circle for detection tasks or another RF pattern for all other tasks). Threshold was reached when at least one spectrum component exceeded a fixed value. Results. The model replicated the observed inefficiency of detecting RF=1, and the hyperacuity at RFs>2. The model also replicated the results of an identification task, where performance decreased at higher frequencies. Model performance is also compared to psychophysical results of masking and increment threshold tasks. Discussion. This represents the first model of the perception of RF patterns. Thus, the model may be relevant to research on shape perception, including face shape, fruit shape, and other curved shapes. This model is compatible with recent data on V4 population coding.
NSERC grant #OP227224 to HRW