A biologically plausible model of human radial frequency perception

Abstract

Several recent studies have used radial frequency patterns to investigate intermediate-level shape perception, a critical precursor to object recognition. Here, we developed the first neural model of RF perception based on known V4 properties that exhibits many of the characteristics of human RF perception. The model is composed of two main parts: (1) recovery of object position using large-scale non-Fourier V4-like concentric units that respond at the center of concentric contour segments across orientations, and (2) curvature detectors that encode local shape information. Each curvature mechanism combines multiplicatively the responses of three oriented filters, the positions and orientation preferences of which determine the curvature mechanism's tuning properties for position, orientation, and degree of curvature. When responding to RF patterns, peak responses occur at points of maximum curvature. Shape is represented as curvature responses as a function of orientation around the object center, and the cross-correlation of that function with a sine wave peaks when the frequency of the sine wave matches the number of peaks in the stimulus. Cross-correlation strength can be used to model human performance. Model and human performance are comparable for detection, identification, and lateral masking tasks. Moreover, the model also shows size invariance of detection performance due to scaling of the curvature mechanisms. The model is then used to make novel predictions.

The curvature calculation mechanism based on simple cell model

“…Among earlier research, Dobbins et al proposed the most promising mechanism: an end-stopping mechanism [1,26,27], well-known for being curvature-selective. However, due to ambiguities in responses obtained using Dobbins et al's method, this paper proposes that it is more accurate to distinguish the curvature and length of line segments in comparison to curvature detectors [10][11][12]28]. In general, cells in V1 had response ambiguities due to the simplicity and small size of their receptive fields; consequently, the subsequent results could only reflect partial features of contours.…”

“…Considering there were far more curves than bars in natural images, and if endstopped cells were capable of measuring curvature, it was unsurprising that end-stopped cells would be almost as common as end-free cells in V1 [4,7,8]. Another two candidates to potentially solve this problem might be (1) an orientation-selective simple cell [9,10], and (2) multiplicative combinations of orientation-selective simple cells [11,12]. This might result in the view that a contour curvature is signaled via the population response of curvature-sensitive neurons tuned to different curvatures for curvature processing [13].…”

Norms are Not the Norm: Testing Theories of Sensory Encoding Using Visual Aftereffects

“…Until now, our knowledge of shape aftereffects had been consistent with a neural substrate in which shapes are represented by the magnitudes and positions of points of maximal local curvature in the retinal image, indicating V4 as a plausible substrate (Bell et al, 2008;Gheorghiu & Kingdom, 2007;Pasupathy & Connor, 2002;Poirier & Wilson, 2006;Suzuki, 2003). The influence of shape constancy operations are not, however, evident until later stages in the visual hierarchy (Janssen et al, 1999;Kourtzi & Kanwisher, 2001;Tanaka, 1996Tanaka, , 2003.…”

“…Shape aftereffects have been linked to V4 activity (Gheorghiu & Kingdom, 2007;Müller, Wilke, & Leopold, 2009;Suzuki, 2003), wherein shape may be encoded by the position and magnitude of points of maximal local curvature (Pasupathy & Connor, 2002). Models of shape perception based on this information can achieve size and position invariance (Poirier & Wilson, 2006), both established properties of the shape aftereffect (Regan & Hamstra, 1992;Suzuki & Cavanagh, 1998). Note, however, that in such models the description of a shape depends solely on the shape of its retinal image.…”

“…This dissociation between retinal image shape and perceived shape is overcome by shape constancy calculations, which enable more accurate shape perception by discounting the retinal image changes that result from viewing objects from different perspectives. These operations are not included in our current understanding of shape aftereffects (Gheorghiu & Kingdom, 2007;Poirier & Wilson, 2006).…”