Reconstruction of CASSI-Raman Images With Machine-Learning
Raman spectroscopy is a well-established method to detect small amounts of potentially dangerous substances. In a Coded Aperture Snapshot Spectral Imaging (CASSI) system spatial and spectral information are mixed resulting in an ensemble of compressed sensing measurements. A reconstruction method is applied to the Compressed Sensing (CS) measurement to reconstruct a hyperspectral cube containing the Raman spectra for the locations in the scene. Traditional reconstruction methods based on regularization such as Total Variation (TV) are time consuming which reduce the number of applications where the technology is applicable. A machine learning reconstruction approach using Convolutional Neural Network (CNN) is presented. The loss function for the CNN is a combination of reconstruction error and re-projection error of the reconstructed Raman spectra. Simulation of CS-measurements of samples containing different chemical substances and different concentration levels are reconstructed with high precision. The reconstruction time using the novel machine learning approach decreases several orders of magnitude.