CNNs with radial basis input function

Abstract

This paper proposes a cellular neural network (CNN) model with radial basis input function (radial basis input CNN) for improving function approximation ability of CNNs. The model can be viewed as a cascade of two units: the first unit is a multi-input, multi-output radial basis function network (RBFN), the second unit is the original CNN model. The weights and centers of the RBFN unit are chosen identical for all RBFN outputs yielding a space-invariant connection weight pattern over the network. With such a weight sharing property, the proposed model becomes a special kind of nonlinear B-template CNN. The ability of the radial basis input CNN model in approximation to functions as its input-(steady state) output mapping is examined on an edge detection task for noisy images. A modified version of the recurrent perceptron learning algorithm (RPLA) is used for the training radial basis input CNN.