Self-Supervised Learning for Raman Spectra Denoising and Peak Deconvolution Under Low SNR

Abstract

Raman spectroscopy is a pivotal analytical technique in chemical physics, molecular biology, and material science, offering a non-destructive fingerprinting capability for molecular identification. However, the practical utility of Raman scattering is frequently impeded by its inherently weak signal intensity, which results in a low Signal-to-Noise Ratio (SNR) when acquisition times are limited or when samples are sensitive to photo-degradation. Traditional computational methods for spectral restoration, such as Savitzky-Golay filtering or wavelet transforms, often necessitate manual parameter tuning and risk distorting peak fidelity, particularly in the preservation of Full Width at Half Maximum (FWHM) values essential for deconvolution. Furthermore, while supervised Deep Learning (DL) models have shown promise, they suffer from a reliance on paired clean-noisy datasets, which are experimentally prohibitive to obtain for complex biological mixtures. This paper presents a novel Self-Supervised Learning (SSL) framework, the Masked Spectral Reconstruction Network (MSR-Net), designed to denoise Raman spectra and facilitate peak deconvolution without requiring ground-truth clean references. By leveraging a masked autoencoding pretext task adapted for 1D correlated signals, the model learns the underlying morphological semantics of Lorentzian and Gaussian peaks while treating stochastic noise as non-reconstructible high-frequency artifacts. We evaluate the approach on both synthetic datasets and real-world mineralogical spectra from the RRUFF database. Experimental results demonstrate that MSR-Net achieves superior SNR improvement and peak position accuracy compared to classical baselines and supervised counterparts trained on limited data.