AI-Interpreted Optical Scattering for Robust and Focal Depth-Aware Imaging

This paper investigates when optical scattering can help rather than hinder imaging. Using No Scattering MNIST and three Scattering MNIST datasets generated under different scattering conditions, the authors analyze speckle patterns with a Variational Autoencoder (VAE), chosen for its interpretable latent space. The abstract reports that scattering improves robustness to spatial pixel loss by distributing information across the observation, and that scattering patterns can encode focal-depth distinctions. The work suggests possible applications in imaging through obstacles and in reconstructing three-dimensional signals, though the current evidence is based on MNIST datasets and the abstract does not provide detailed quantitative results.