An analog network for continuous-time segmentation

摘要

A common goal in computer vision is to segment scenes into different objects sharing common properties such as depth, motion, or image intensity. A segmentation algorithm has been developed utilizing an absolute-value smoothness penalty instead of the more common quadratic regularizer. This functional imposes a piece-wide constant constraint on the segmented data. Since the minimized energy is guaranteed to be convex, there are no problems with local minima, and no complex continuation methods are necessary to find the unique global minimum. This is in sharp contrast to previous software and hardware solutions to this problem. The energy minimized can be interpreted as the generalized power (or co-content) of a nonlinear resistive network. The network is called thetiny-tanh network since the I-V characteristic of the nonlinear resistor must be an extremely narrow-width hyperbolic-tangent function. This network has been demonstrated for 1-D step-edges with analog CMOS hardware and for a 2-D stereo algorithm in simulations.