Classification of blurred textures using multilayer neural network based on multi-valued neurons

In this paper, we consider the problem of blurred texture classification using a multilayer neural network based on multi-valued neurons (MLMVN). We use the frequency domain as a feature space. The low frequency part of the Fourier phase spectrum of a blurred image remains almost unaffected by blur. This means that phases corresponding to the lowest frequencies can be used as features for classification. MLMVN is the most suitable machine learning tool for solving the problem, since it uses phases as inputs. MLMVN is based on multi-valued neurons whose inputs and output are located on the unit circle and therefore they are determined exactly by phases. This determines a very important ability of MLMVN and MVN to treat phases properly We employ in this paper a slightly modified learning MLMVN rule and a modified learning strategy, which extends margins between classes´ representatives used for the learning and the borders of classes. This approach makes it possible to classify with 100% accuracy even such heavily blurred textures where visual analysis and classification are not possible at all.