Spectral Raman unmixing from CASSI system compressive measurements

Raman spectroscopy is a non-invasive technique that characterizes the biochemical components of a sample using the Raman scattering effect. A Raman data cube is a three-dimensional set comprising two spatial and one spectral dimensions. Typically, thousand bands per pixel are captured which considerably increases acquisition times. Furthermore, many pixels capture information from a mixture of several components, thus decreasing the accuracy of substances classification. For this reason, unmixing methods emerge as a solution to this problem. Traditionally, the SUnSAL algorithm is used to solve the spectral mixing problem on linear mixing assumption [1]. Additionally, the recent theory of compressive sensing (CS) provides a solution to the excessive size of the data cube and large acquisition times. An efficient CS optical implementation is the coded aperture snapshot spectral imaging (CASSI) system which capture spatial and spectral information of scene. This paper presents the spectral unmixing process on CASSI measurements of Raman spectral images. The SUnSAL algorithm is modified to be used with CASSI images and a signature re-scaling step is added to take in account the variability of signatures by different capture set-up. Extensive simulations with real and synthetic Raman images show the accuracy of this algorithm.