Neurobiologically Inspired Geometric Diffusion for Target Recognition
Abstract
We address the target recognition problem by focusing on intermediate-level vision. Early biological vision extracts edges and contours of various lengths. High-level recognition is either view or template-based, which is fragile with respect to lighting, size, or clutter; or medial-axis-based, which requires a perfect bounding contour. Diffusion processes are central to neurobiology, and we have discovered how to use them to bridge the gap between (local) edges and (global) descriptions for matching. We have proved that the equilibria of these distributions signals information from the distance map that underlies medial axis computations. This equilibrium distribution therefore makes explicit global pattern from local features and can be used for matching. We have shown how this provides a novel solution to detecting airports and other "complex features" in imagery, and how it suggests a novel solution to the border-ownership problem in neuroscience.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 12, 2012
- Accession Number
- ADA577270
Entities
People
- Steven W. Zucker
Organizations
- Yale University