Application of Biorthogonal Filter Functions to Pattern Recognition and Feature Extraction

摘要

A mathematical method is developed for generating a set of filter functions from a given set of signatures. The filter function set of functions is biorthogonal to the set of signature functions; therefore, any one filter function gives a perfect response to one signature, and a response to all other given signatures is completely suppressed. The method can be used to decompose superpositions of signatures as well as for improving separation of measured parameters for pattern recognition. It can also be used to suppress interferences from the background when it is included in the given set of signatures. A method of adding new filter functions to an existing set without complete recomputation (adaptive learning) is discussed. (Author)