Trainable Diffractive Surfaces for Spectral Encoding of Spatial Information

We demonstrate a deep-learning based single-pixel optical machine vision framework, where multiple diffractive surfaces are used to transform and encode the spatial information of objects into the power spectrum of the diffracted light. Specifically, by predetermining a set of wavelengths each representing a data class, we trained diffractive surfaces to maximize the power of the diffracted wavelength corresponding to the correct data class to perform all-optical object classification through a single-pixel detector. Using a plasmonic nanoantenna-based spectroscopic detector and 3D-printed diffractive layers, we experimentally validated this design by successfully classifying handwritten-digits using a snap-shot broadband illumination. Further, we combined this all-optical spectral encoding scheme with a separately trained shallow artificial neural network to improve the inference accuracy through a feedback between the optical and electronic networks. The same electronic network was also used to reconstruct the images of the input objects solely based on the power of target spectral components, demonstrating the success of our framework as a resource-efficient, data-specific machine vision platform.