DF-LDA tree: a nonlinear multilevel classifier for pattern recognition

This article presents a new nonlinear classifier by arranging linear classifiers in a tree structure. The proposed classifier, called the direct fractional-step linear discriminant (DF-LDA) tree, adopts a tree structure containing a DF-LDA at each node. The structure of the tree classifier evolves as the training proceeds, so there is no need to decide any parameters as a priori. Due to the many DF-LDAs arranged in the tree structure, classification performance of the proposed classifier is improved over single-shot DF-LDA. The proposed DF-LDA tree is tested on various synthetic and real datasets. Experimental results show that the proposed classifier leads to very satisfactory results in terms of classification accuracy.