Pattern recognition and chemometrics for spectral classification

摘要

The broader defintion of chemometrids includes methods such as pattern recognition (PR) and signal/image processing for nonivasive analysis and interpretation of data. These methods are among the most pwoerful tools currently available for nonivasively examining spectroscopic and other chemical data. Using spectral data, these systems have found a variety of applications employing anlaytical techniques for gas chromatography, fluorescence IR or NMR spectroscopy, etc. An advantage of PR approaches is that they make no a priori assumption regarding the structure of the spectra. However, a majority of these systems rely on human judgment for parameter selection and classification of spectra. Generally a spectral pattern recognition (SPR) problem is considered as a group of several subproblems. We considered a SPR problem as a group of five subproblems: spectra acquisition, feature extraction, feature selection, spectra organization, and spectra chassification. One of the basic issues in PR approaches is to determine and measure the discriminatory features useful for successful classification. A spectral pattern classification system, combining spectral feature extraction and selectrion, and decision-theoretic approaches, is developed. It is shown how such a system can be used for analysis of large data analysis, warehousing, and interpretation.