Peak-aware adaptive denoising for Raman spectroscopy based on machine learning approach

Raman spectroscopy can be effectively used for detection and analysis of chemical agents that are serious threats in modern warfare, but the detection and analysis performance is prone to deterioration due to noise. The existing denoising technique has limitations that there is no criterion for selecting the window length and that the filtering distorts the peaks, key features for Raman spectral data analysis. To overcome such limitations, in this paper, we propose the peak-aware adaptive denoising for Raman spectroscopy based on machine learning approach. The proposed technique utilizes the information of detected peaks to eliminate noise effectively using different window values optimal for each region in the Raman spectrum while preserving the shape of peaks. We conducted the various analyses and experiments, and the proposed technique showed a 28% lower Euclidean distance and a 48% lower Frechet inception distance compared to the existing technique, meaning the proposed technique outperformed the existing one. In this study, we propose the novel peak-aware adaptive denoising technique tailored for Raman spectroscopy, aiming to overcome the limitations inherent in current methodologies that lack predefined criteria for selecting window length and may distort spectral peaks. Our proposed technique leverages machine learning and harnesses information derived from detected peaks to systematically eliminate noise while applying the optimal window length for each spectrum region, effectively reducing noise and adeptly preserving peak shapes. Comparative analyses show that our technique demonstrates a noteworthy 28% reduction in Euclidean distance and a substantial 48% decrease in Frechet inception distance when contrasted with existing denoising techniques.image