A theoretical analysis on nearest feature space classifier

Classifier design is an important issue in pattern recognition. Nearest Feature Line (NFL) classifier had been proposed to enhance the prototype-representing capacity of nearest neighbor methods. Nearest Feature Space (NFS) is further proposed as a generalization of NFL. In this paper we give a formal definition and theoretical analysis on Feature Space. First, the dimensionality and coordinate origin problems in classical NFS are presented. Then, a novel NFS classifier is designed to solve the above-mentioned problems. For contrastive analysis, a case study on image recognition is carried out by using ORL database. Experimental results indicate that the proposed NFS classifier offers a better recognition performance than classical NFS, which practically proves that the proposed NFS accommodates an outstanding prototype-representing capacity.