Denoising Raman spectra using a single layer convolutional model trained on simulated data

Abstract

Raman spectroscopy is a powerful means of revealing chemical and structural information about a sample and acquiring chemically specific images. Such images often suffer from low signal to noise ratios (SNR). In this report, a novel way to improve the SNR using machine learning tools based on simulated data. The proposed approach offers an alternative to time consuming acquisition and labeling of large data sets and can be readily applied to unknown systems. Here, the efficacy of a single layer denoising network trained only on simulated data was evaluated, and it was found that the proposed model was able to provide a substantial improvement in SNR.

Artificial Intelligence for Surface‐Enhanced Raman Spectroscopy

“…Gil et al have proven that a CNN with only one convolutional layer with sigmoid activation function can outperform a single implementation or even 25 cascade implementations of SG filter. [ 110 ] Other CNN‐based models have also demonstrated the joint function of denoising and baseline correction with full noise deplete and least peak reduction. [ 108,109,111,112 269 ] Unsupervised algorithms can also be leveraged in denoising, specifically without the requirement of corresponding denoised spectra.…”