Variational auto-encoder based deep compressed sensing on Raman spectroscopy

Raman spectroscopy is a non-intrusive and rapid profiling technology. Its hardware depends on an ADC (Analog to Digital Converter) unit to convert photons into electronic signals. Previous studies have proved that ADC can benefit from CS (Compressed Sensing), allowing a much lower sampling rate and faster sensing speed. Compared to the general-purposed non-adaptive CS transforms (e.g., discrete cosine / Fourier transforms), this study tries to design an adaptive CS based on the recent deep learning technology. A variational auto-encoder (VAE) model is developed and pre-trained on the domain spectroscopic profiling data. The VAE’s decoder part ( G ) is a domain-specific generative model that maps the latent embedding ( z ) to the original feature space ( x ). In the CS scenario, the decoder G acts as the transform basis ( ψ ). In a Raman spectroscopic case study of distinguishing between 5-year and 26-year vintage liquors, the proposed VAE-CS can achieve 100% accuracy at only 1% sampling percentage. Meanwhile, non-adaptive CS using DCT (Discrete Cosine Transform) and DFT (Discrete Fourier Transform) needs 10% to achieve a comparable classification accuracy. The result shows that VAE-CS is much more efficient than the non-adaptive counterpart. The result shows that our method is ten times more efficient than the non-adaptive counterpart. This means the pre-trained VAE model has learned useful domain-specific knowledge and the decoder G(z) can leverage this knowledge for high-quality signal reconstruction.