Training pattern replication and weighted class allocation in artificial neural network classification

Abstract

In some image classifications the importance of classes varies, and it is desirable to weight allocation to selected classes. Often the desire is to weight allocation in favour of classes that are abundant in the area represented by an image at the expense of the less abundant classes. If there is prior knowledge on the distribution of class occurrence, this weighting can be achieved with widely used statistical classifiers by setting appropriate a prioriprobabilities of class membership. With an artificial neural network, the incorporation of prior knowledge is more problematic. An approach to weight class allocation in an artificial neural network classification by replicating selected training patterns is presented. This investigation focuses on a series of classifications in which some classes were more abundant than others, but the same number of training cases were available for each class. By replicating the training patterns of abundant classes the representation of the abundant classes in the training set is increased, reflecting more closely the relative abundance of the classes in an image. Significant increases in classification accuracy were obtained by replicating the training patterns of abundant classes. Furthermore, in comparison against a discriminant analysis for the classification of synthetic aperture radar imagery, the results showed that training pattern replication could be used to weight class allocation with an effect similar to that of incorporating a prioriprobabilities of class membership into the discriminant analysis, and resulted in a significant 20.88%, increase in classification accuracy. This increase in classification accuracy was obtained without any new information, but was the result of making fuller use of what was available.