Rapid material identification via low-resolution Raman spectroscopy and deep convolutional neural network

Abstract

Raman spectroscopy is a vital technique being able to detect and identify molecular information with advantages of being fast and non-invasive. This technique also enables numbers of potential applications, including forensic drugs detector, explosive detection, and biomedical analysis. In this work, we investigated the identification performance of a custom-made low-resolution Raman system equipped with machine learning capability to classify various types of materials. Here, a relatively broadband laser diode with center wavelength of 808 nm was used for Raman excitation. An off-axis parabolic mirror with through hole was used in place of a beamspiltter for sample excitation, as well as collection, and collimation of scattered light from long working distance of 50 mm. The signal was filtered and delivered to a cooled spectrometer via an optical fiber for spectra measurements. Raman spectra of test samples were on the range of 100-2000 cm-1 with 7.65 cm-1 data steps. For spectral analysis, a convolutional neural network (CNN) was implemented as classification algorithm with feature extraction from multiple layers together with error-back propagation, which displayed the performance in term of accuracy. It was found that with only three sets of convolution layers up to 96.7% testing performance can be achieved even with low spectral resolution input.