Assessments of neural network output codings for classification of multispectral images using Hamming distance measure

Abstract

The influences of different output codings on the performances of a back-propagation neural network for classification of multispectral images are investigated. The assessments of the output codings are based on the convergence ability, training and classification performances. Effects of the mapping permutation of output states to the information classes are discussed. Results obtained for all the possible permutations, of each coding scheme, are presented in terms of the statistical mean, variance, maximum and minimum of the performance features. Hamming distance measure is introduced as a tool to access the credibility of the coding schemes. Results obtained show that output coding schemes with equal Hamming distance between the output states have better generalization properties and performances obtained for different output mapping permutations are consistently high.