Nonlinear analog networks for image smoothing and segmentation

Image smoothing and segmentation algorithms are frequently formulated as optimization problems. Linear and nonlinear (reciprocal)resistive networks have solutions characterized by an extremum principle. Thus, appropriately designed networks canautomatically solve certain smoothing and segmentation problems in robot vision. This paper considers switched linear resistive networks and nonlinear resistive networks for such tasks. Following 1] the latter network type is derived from the former via an intermediate stochastic formulation, and a new result relating the solution sets of the two is given for the zero temperature limit. We then present simulation studies of several continuation methods that can be gracefully implemented in analog VLSI and that seem to give good results for these nonconvex optimization problems.