BayesOpGAN: A Bayesian-Optimized GAN Framework with Fourier-Based Evaluation for Quality-Controlled Raman Spectral Data Augmentation

Deep learning models for Raman spectral classification often overfit and generalize poorly when trained on small data sets. To overcome this limitation, we present BayesOpGAN, a generative adversarial network incorporating a Bayesian-optimized loss (BayesOpLoss) and smooth upsampling modules to produce fidelity Raman spectra from fewer than 30 samples. The quality of generated spectra is quantitatively assessed using Fourier Distance, a frequency-domain metric designed for one-dimensional signals. The framework is validated on a hybrid-quality data set from the RRUFF Raman database, where BayesOpGAN-augmented data enhance the performance of a ResNet-50 classifier. Test accuracy increases from 83.9% to 91.0%, while excessive generated data reduce performance, indicating an optimal augmentation range. Overall, BayesOpGAN provides a reliable and scalable solution for small-sample spectral data augmentation, where Fourier Distance ensures objective fidelity evaluation and generalizes well to other one-dimensional spectroscopies and biomedical sensing.