Abstract:

This paper presents a new classification technique. The proposed method is named the Local Subspace Classifier (LSC) which indicates the kinship of the algorithm to the subspace classification methods. On the other hand, the Local Subspace Classifier is an heir of pure prototype classification methods like the k-NN rule. Therefore, it is argued that, in a way, LSC fills the gap between the subspace and prototype methods of classification. From the domain of the prototype-based classifiers, LSC brings the benefits related to the local nature of the classification, while simultaneously utilizing the capability of the subspace classifiers to produce generalizations from the training samples. A further enhancement of the LSC principle named the Convex Local Subspace Classifier (LSC+) and the application of the LSC principle to SOM are presented. Experiments with handwritten digit data demonstrate the good classification accuracy obtainable with the LSC and LSC+ classifiers.