Scale-based normalization of spectral data.
Abstract
Classification of data that arise as signals or images often requires a standardization step so that information extracted from biologically equivalent signals can be quantified for comparison across classes. Differences in global trend, total energy, high-frequency noise and/or local background can arise from variabilities due to instrumentation or conditions during data collection. This article considers some common ways in which such variation is adjusted for and introduces a generalization of the popular "standard normal variate" transformation. Based on a wavelet decomposition this generalization provides increased flexibility for normalizing spectral data affected by local background noise. Examples from three types of spectroscopy data illustrate the method and its properties.