Space-Based Computational Hyperspectral Machine Vision using Compressed Sensing Neural Networks

In this STTR Phase I proposal, Kent Optronics (KOI) together with its partner, Rice University, propose to develop novel deep learning algorithms to perform machine vision tasks such as target recognition and tracking utilizing the direct measurements from a compressive hyperspectral imaging system. By skipping the hypercube reconstruction, this combination of hardware and software will allow real-time, actionable reaction to the incoming datastream. The proposed space-qualified computational sensor is a hyperspectral imager based on the principle leveraging on a structured illuminator. In combination with the sparse recovery algorithms the sensor can efficiently recover the volume density of a participating medium which is described by volume densities rather than boundary surfaces, e.g. translucent objects, smoke, clouds, mixing fluids, and biological tissues. In Phase I, a thorough trade analysis and model validation test on both a manifold secant learning algorithm will be compared with a novel dynamic multi-rate compressive neural network approach in simulation. In Phase II, both of these approaches will be incorporated, tested and qualified in real-world compressive hyperspectral imaging hardware