Superresolution for Ultrasonic Imaging in Air Using Neural Networks.

In this paper we study ultrasonic imaging in air using an array of transducers. We describe a superresolution technique that uses the fact that most surfaces act as perfect reflectors to ultrasonic pulses in air to generate accurate surface maps for object identification. The technique involves the minimization of a quadratic objective function subject to a nonlinear (quadratic) equality constraint. We show that this minimization can be accomplished by a two step quadratic programming method, which, although not practical on a general purpose computer, can operate in real time on a pair of neural networks. Results demonstrate that the technique generates accurate surface maps even with low receive signal-to-noise ratios.