The Nearest Feature Midpoint - A Novel Approach for Pattern Classification

In this paper, we propose a method, called the nearest feature midpoint (NFM), for pattern classification. Any pair of feature points of the same class is generalized by the feature midpoint (FM) between them. Hence the representational capacity of available prototypes can be expanded. The classification is determined by the nearest distance from the query feature point to each FM. This paper compares the NFM classifier against the nearest feature line (NFL) classifier, which has reported successes in various applications. In the NFL, any pair of feature points of the same class is generalized by the feature line (FL) passing through them, and the classification is evaluated on the nearest distance from the query feature point to each FL. The NFM can be considered to be the refinement of the NFL. A theoretical proof is provided in this paper to show that for the n-dimensional Gaussian distribution, the classification based on the NFM distance metric will achieve the least error probability as compared to those based on any other points on the feature lines. Furthermore, a theoretical investigation is provided that under certain assumption the NFL is approximately equivalent to the NFM when the dimension of the feature space is high. The experimental evaluations on both simulated and real-life benchmark data concur with all the theoretical investigations, as well as indicate that the NFM is effective for the classification of the data with a Gaussian distribution or with a distribution that can be reasonably approximated by a Gaussian.