Pattern Recognition, 4th Edition

5.8. Optimal Feature Generation

So far, the class separability measuring criteria have been used in a rather “passive” way, that is, to measure the classification effectiveness of features generated in some way. In this section we will employ these measuring criteria in an “active” manner, as an integral part of the feature generation process itself. From this point of view, this section can be considered as a bridge between this chapter and the following one. The method goes back to the pioneering work of Fisher ([Fish 36]) on linear discrimination, and it is also known as linear discriminant analysis (LDA). We will first focus on the simplest form of the method in order to get a better feeling and physical understanding of its basic rationale. ...