Enhancement of Shape-from-Texture Perception by Spatial Frequency Processing

Ko Sakai¹, Shih-Cheng Yen², and Leif H. Finkel²

¹Institute of Physical and Chemical Research (RIKEN), FRP, Laboratory for Neural Modeling, Wako, Japan
²Department of Bioengineering, University of Pennsylvania, Philadelphia, PA

Abstract

Purpose. Changes in the slant and tilt of a textured surface are reflected in the spatial frequency domain as affine transformations of the frequency spectrum. We have previously presented evidence (Sakai and Finkel, J. Opt. Soc. Amer., 1995) that perception of surface curvature is correlated with changes in the average peak frequency (APF), the peak frequency at each image location averaged over a spatial neighborhood. These results suggested that the visual system may simply track APF as opposed to computing changes across the entire spectrum. We present a demonstration that APF is also correlated with estimation of depth for planar surfaces viewed in perspective projection.

Methods. A series of textured surfaces, both artificially rendered and real images, were used as inputs to a network model based on extraction of APF.

Results. Depth estimates of the network corresponded with human judgments. The dependence of APF solely on the peak spectral frequency confers a resistance to image noise. Simulations show a graceful degradation of curvature estimates as s/n decreases.

Conclusions. These results suggest that the contribution of texture-based processes in shape discrimination and image segmentation may be strengthened by preprocessing to emphasize strong components in the frequency spectrum. Such a process could be carried out by gain-control mechanisms.

Supported by ONR N00014-93-1-0242 and The Whitaker Foundation