Cauchy filters versus neural networks when applied for reconstruction of absorption spectra

The computer-based interpretation of data {y/spl tilde//sub n//sup Tr/}, acquired by means of an optical spectrometer, is aimed at identification of the main components of an analyzed substance. The first step of interpretation consists in estimation of its spectrum {x/sub n//sup Tr/} using an operator of (generalized) deconvolution {x/spl circ//sub n//sup Tr/} = R[{y/spl tilde//sub n//sup Tr/}; p/sub R/], where p/sub R/ is a vector of parameters to be estimated during calibration of the spectrometer. Several new structures of this operator, based on combinations of a nonlinear transformation or the Cauchy filter and of an radial-basis-function-type neural network, are proposed and studied in this paper using both synthetic and real-world spectrophotometric data. Their superiority over existing operators for spectrum reconstruction is demonstrated with respect to uncertainty of reconstruction.