Deep Residual Neural Networks For Robust Denoising In Raman Spectroscopy

Raman spectroscopy is a key tool for material analysis, but its accuracy is often hindered by noise and baseline distortions. This paper presents a robust denoising method using a parallel deep residual neural network architecture based on DnCNN, designed for one-dimensional spectral data. The model learns noise patterns through multiple convolutional branches, enabling effective denoising without assumptions about the signal origin. We evaluate several pre-processing techniques, with minimum-shift normalization proving most effective in preserving spectral features. Trained on datasets with varying noise levels, the network achieves high peak detection accuracy and low error rates, outperforming traditional and recent methods. This approach enhances the reliability of Raman analysis and demonstrates the potential of AI-driven models in spectroscopy and time-series signal processing.