Toward a unified design of pattern recognizers

Finding salient features is one of the most important steps in designing a pattern recognizer. Although this issue has long been studied from the early days of research on pattern recognition, no study has yet provided a well-formalized satisfying solution which is able to be easily used in practical conditions. In this paper, we discuss in detail a new family of discriminative feature design methods, namely, discriminative feature extraction, the minimum error learning subspace method and discriminative metric design. Tracing the development of these methods clarifies a desirable research direction toward a solution to the problem.