Analog Models for Early Vision
Author: Harris, John Gregory
Year: 1991
Degree: Dissertation (Ph.D.)
Advisor: Koch, Christof
Committee Members: Koch, Christof; Barr, Alan H.; Mead, Carver; Van Essen, David C.; Goodman, Rodney M.
Option: Computation and Neural Systems
DOI: 10.7907/CTH4-5074
Abstract
Analog models provide a novel framework for understanding and developing algorithms for computer vision. This thesis introduces several extensions to well-known resistive network techniques for solving early vision problems. First, constraint boxes are developed as a general methodology for mapping regularization-based algorithms onto stable analog hardware. These multiterminal resistor systems solve low-level vision problems by minimizing a global Lyapunov energy. Second, a circuit element called the resistive fuse is introduced to extend these networks for discontinuity detection. This is the first hardware circuit that explicitly implements line-process discontinuities. Since resistive fuse networks must minimize a non-convex energy function that may contain local minima, complex annealing or continuation methods are necessary for adequate solutions of the problem. Third, the tiny-tanh network is proposed as a new mechanism for discontinuity detection that is not plagued by problems with local minima. A piece-wise constant segmentation is performed through minimization of a convex Lyapunov energy.
Files
- Harris_jg_1991.pdf (application/pdf)