Comparing AutoEncoder Variants for Real-Time Denoising of Hyperspectral X-Ray

Hyperspectral X-ray analysis is used in many industrial pipelines, from quality control to detection of low-density contaminants in food. Unfortunately, the signal acquired by X-ray sensors is often affected by a great amount of noise. This hinders the performance of most of the applications building on top of these acquisitions (e.g., detection of food contaminants). Therefore, a good denoising pipeline is necessary. This article proposes a comparison between three different AutoEncoder variants: the Variational AutoEncoder, the Augmented AutoEncoder, and a plain vanilla AutoEncoder. All the networks are trained in an unsupervised fashion to denoise a given noisy spectrum. Focusing on the specific application of recognizing possible food contaminants, we force the latent space of the networks to have just two parameters, as suggested by the physical law of Lambert-Beer. We validate our experiments on a synthetic dataset composed of roughly 15 million spectra. Results suggest that the Augmented AutoEncoder is the best network configuration for this task, showing excellent performance without suffering from the nondeterministic behavior of the Variational AutoEncoder.