Research on Raman Spectral Denoising and Baseline Correction Method Based on Dynamic Mask Partition U-Net

Raman spectroscopy, as a core technique for qualitative analysis of material structure and quantitative analysis of composition, is widely applied in fields such as materials science, biomedicine, and environmental monitoring. However, Raman spectral signals generally suffer from weak intensity and low signal-to-noise ratio, and are susceptible to interference from fluorescence background and instrument electronic noise, leading to baseline drift, blurred or even disappearing characteristic peaks. Traditional denoising methods mostly adopt a single smoothing strategy or global processing mode, making it difficult to adapt to the differentiated denoising requirements of different spectral regions, and often resulting in problems such as peak distortion or insufficient noise suppression. To this end, this study proposes a joint preprocessing method based on dual-Mask peak-Aware U-Net joint denoising and baseline correction (DMPA-UNet). First, a lightweight convolutional network is used to generate an adaptive mask, dynamically dividing the spectrum into peak regions and non-peak regions; then, a dual-branch U-Net is adopted to process the two regions differently, and mask-weighted fusion is used to obtain the denoised spectrum; finally, an independent U-Net is used to perform global baseline estimation and removal, achieving collaborative optimization of denoising and baseline correction. Comparative experiments between this method and other methods, such as SG (Savitzky–Golay) filtering, median filtering, and CNNs (Convolutional Neural Networks), show that, compared with the original noisy spectrum, this method reduces MSE by approximately 80.9%, improves SNR (signal-to-noise ratio) by approximately 116.0%, and improves PSNR (peak signal-to-noise ratio) by approximately 38.9%; the KNN (K-Nearest Neighbor) classification accuracy can reach 84.5%, an improvement of 69.5% over the original noisy spectrum. This method balances peak preservation and noise suppression, providing an effective preprocessing scheme for qualitative analysis of Raman spectra.