CNN based self-adjusting nonlinear filters

In this paper, we discuss CNN based adaptive nonlinear filters derived from robust statistic and geometry-driven diffusion paradigms. The base models of both approaches are defined as difference controlled nonlinear (DCN) CNN templates while the self-adjusting property is ensured by simple analogic (analog and logic) CNN algorithms. The proposed methods provide a practical framework for VLSI implementation, since all nonlinear cell interactions of the CNN architecture are deduced to two fundamental nonlinearities, to a sigmoid-type and a radial basis function. These nonlinear characteristics in DCN templates can be approximated by simple piecewise-linear functions of the difference voltage of neighboring cells. The simplification makes possible to convert all space invariant nonlinear templates of this study to a standard instruction set of the CNN Universal Machine, where each instruction is coded by at most 10 analog numbers. Through examples it is demonstrated, that such CNN based adaptive nonlinear filters have excellent performance in filtering both the impulsive and Gaussian noise while preserving the image structure.