Deep metric learning framework combined with Gramian angular difference field image generation for Raman spectra...

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Portable Raman spectroscopy is crucial for rapid material identification.
• Challenges include spectral resolution differences and model training efficiency.
• Existing classification models struggle with large numbers of substance classes.

Purpose of the Study:

• To develop a fast and accurate Raman spectral classification method.
• To address spectral dimension disparities across different Raman devices.
• To improve the efficiency of developing substance identification devices.

Main Methods:

• Utilized deep metric learning networks combined with Gramian angular difference field (GADF) image generation.
• Converted Raman spectra to GADF images for uniform resolution.
• Trained the network on 450 mineral classes from the RRUFF database.

Main Results:

• Achieved 98.05% accuracy for 260-class mineral classification and 90.13% for 8-class pathogenic bacteria classification.
• Demonstrated superior performance over mainstream machine learning models.
• Attained 99.14% accuracy in identifying 32 classes of chemical substances using a handheld spectrometer.

Conclusions:

• The proposed method significantly improves Raman spectral classification efficiency and accuracy.
• The approach is suitable for embedded systems in portable Raman spectrometers.
• Enables widespread application in unknown substance identification tasks.