Fisher subspace tree classifier based on neural networks

Abstract

This paper proposes a multi-neural network classification based on fisher transformation. The new method improves HDR[1] (Hierarchical discriminate regression) method proposed in 2000, which can classify the training set from coarse to fine by non-linear dynamic clustering for high-dimension data. In proposed method a fisher subspace replaces K-L subspace of HDR that simplifies the Hierarchical tree. Simulation results show that our method is better than HDR on recognition ratio and time cost. © Springer-Verlag Berlin Heidelberg 2005.