Design of optimal stack filters: a neural net approach with BP algorithm

Abstract

It is well-known that the design of optimal stack filters has been restricted seriously by the filter's size. The maximum size that can now be handled with the developed techniques is 18 which cannot even cover fully a 5/spl times/5 square mask. In this paper, we present a neural network approach to the optimal design of stack filters where we treat each minimum (MIN) or maximum (MAX) operation as a neuron. In this way, we can design the positive Boolean function (PBF) directly, thus avoiding the determination of the whole Karnaugh map (which may have a prohibitively large size) as required in the conventional methods. The design of an optimal filter is accomplished by using the backpropagation (BP) algorithm. Some specific characteristics concerning the design process are addressed in details, accompanied by a few design examples so as to justify the proposed method.