Distorted character recognition by an incompatible single-layer dipole neural network

Abstract

This paper solves two problems: the first problem is devoted to the recognition of distorted symbolic images by a single-layer incompatible dipole neural network, and the second - the optimization of computing resources in the recognition of distorted symbolic images. In particular, the architecture of an incompatible single-layer network with dipole neurons is proposed. Incompatibility of synaptic connections between neurons is based on the fact that significant interaction between dipole neurons exists in their immediate environment. Synaptic connections between dipole neurons are taken into account only between the nearest neighboring neurons, because the synaptic tensor λij between the i -th and j -th dipole neurons is inversely proportional to the distance rij between neighboring i -th and j -th dipole neurons, therefore λij+1<<λij . An algorithm for recognizing incoming distorted symbolic images using an incompatible dipole neural network has been developed and implemented in the Matlab application system. It is shown that for the recognition of input symbol images by an incompatible dipole neural network the computational resource time is shorter compared to a fully connected neural network by n(n+1)/4 times ( n is the number of pixels in columns and rows, respectively, used for encoding of input images). Numerical experiments have shown that the computational time to recognize 0,4n2 distorted characters, which is described by a 5×5 matrix, is 7,5 times less than the recognition time of a fully connected neural network.