A sparse regularized model for Raman spectral analysis

Raman spectroscopy has for a long time performed as a common analytical technique in spectroscopic applications. A Raman spectrum depends upon how efficiently a molecule scatters the incident light (electron rich molecules often produce strong signals) which results in difficulties for relating the spectrum to the absolute amounts of present substances. The spectrum is however a stable and accurate representation of the sample measured especially considering that each molecule is associated with a unique spectrum. State-of-the-art spectroscopic calibration methods include the principal component regression (PCR) and partial least squares regression (PLSR) methods which have been proved to be efficient regression methods to realise the quantitative analysis of Raman spectrum. In this paper we consider the problem of Raman spectra deconvolution to analyse the sample composition, as well as possible unknown substances. In particular, we propose a sparse regularized model as a complement to traditional regression methods by leveraging the components sparsity compared to the whole chemical library and the spectra sparsity, given that the chemical fingerprint of each spectrum is mainly determined by the peaks. Experimental results illustrate the effectiveness of this sparse regularized model.