A three-stage deep learning-based training frame for spectra baseline correction.

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Baseline drift in spectral data significantly impacts measurement accuracy and quantitative analysis.
• Deep learning offers powerful solutions for spectral baseline correction but is hindered by large paired data requirements and data acquisition challenges.

Purpose of the Study:

• To develop a robust deep learning method for spectral baseline correction that overcomes data limitations.
• To improve the performance of U-net architectures for spectral data processing.
• To reduce the reliance on extensive paired spectral datasets for training.

Main Methods:

• A novel Learned Feature Fusion (LFF) module was designed to enhance U-net performance by adaptively integrating multi-scale features.
• A three-stage training framework was proposed: Stage 1 utilized airPLS for initial data refinement, Stage 2 employed synthetic spectra generation, and Stage 3 used contrastive learning to bridge synthetic and real spectral data gaps.

Main Results:

• The proposed LFF module significantly improved U-net's capability in handling spectral baseline correction.
• The three-stage training framework effectively reduced the dependency on large, paired spectral datasets.
• Experimental results demonstrated the method's efficacy as a powerful tool for baseline correction.

Conclusions:

• The developed deep learning method offers an effective solution for spectral baseline correction, addressing key data limitations.
• The approach shows significant potential for improving spectral quantitative analysis and related applications.