Optimizing spatial filters for single-trial EEG classification via a discriminant extension to CSP: the Fisher criterion.

摘要

In this article, a new spatial filtering approach, called discriminant common spatial patterns (dCSP), is proposed for single-trial EEG classification. Unlike the conventional common spatial patterns (CSP) that is substantially a subspace decomposition technique, dCSP is intently designed for discriminant purpose. The basic idea of dCSP is to construct a Fisher-like criterion that extracts both between-class and within-class discriminant information. The classical CSP only considers separating class means, i.e., between-class scatter, as well as possible. In contrast, dCSP aims to maximize between-class scatter and meanwhile minimize within-class scatter. Computationally, dCSP is formulated as a generalized eigenvalue problem. Experiments on real EEG classification show the effectiveness of the proposed method.