A generic fuzzy neuron and its application to motion estimation

© Springer-Vertag Berlin Heidelberg 1996. The advantages of fuzzy sets and neural networks in emulating the human brain capabilities motivated the development of fuzzy neural networks. Various models of fuzzy neurons have been proposed as the basic element of fuzzy neural networks. In this paper, we introduce a generic fuzzy neuron as an extension of existing fuzzy neuron models. In our model, all the states of activity are given in terms of fuzzy sets with relative grades of membership distributed over the interval [0, 1]. The inputs and outputs are fuzzy sets over different universes of discourse. The connection, aggregation, and activation functions, which determine the operation of the neuron, are fuzzy relations. When the inputs to a function are fuzzy sets over the same universe of discourse, the function can be any fuzzy operation in class of triangular norms or triangular conorms. To evaluate the operation of the fuzzy neuron, a fuzzy neural network architecture based on the generic fuzzy neuron has been developed for motion estimation. The five-layer feed forward fuzzy neural network emulates a fuzzy motion estimation algorithm. Seven simplified versions of fuzzy neurons are defined and utilized in the fuzzy neural network. The results of simulations on thousands of 64×64, 6-bit synthetic image frames containing moving objects under different conditions are reported.