Baseline correction using DU-GridNet for Raman spectra

Raman spectroscopy is a pivotal tool for molecular and material characterization; however, fluorescence backgrounds often severely distort measured signals and hinder downstream analytical tasks. Conventional baseline correction methods typically rely on iterative optimization with manual parameter tuning, rendering end-to-end automated processing difficult. Although deep learning frameworks have emerged to automate background removal, existing models often struggle to balance the preservation of sharp, localized Raman peaks with the accurate modeling of broad, slowly varying baselines. To overcome this limitation, we introduce DU-GridNet, a compact one-dimensional deep network utilizing a multi-resolution grid topology to automate Raman baseline correction. Through multi-scale feature propagation via interconnected sampling blocks, the model simultaneously isolates sharp Raman peaks and fits broad fluorescence backgrounds without manual intervention. Benchmark evaluations demonstrate that the proposed model achieves superior baseline estimation accuracy compared to existing methods. Furthermore, validation on empirical Raman datasets confirms its robust capacity to eliminate diverse fluorescence backgrounds while preserving critical spectral fidelity.