Project

DESCRIPTION:

Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) is a key technique of remote-sensing image recognition, which can be supported by deep neural networks. We aim to develop (1) advanced graph machine learning approaches to enable SAR-based marine applications, (2) hardware acceleration for real-time AI-enabled decision-making, (3) adversarial attacks on machine learning models for SAR-based applications, (4) explainable AI for ML model-based decision making, and (5) high-performance machine learning system based on FPGAs and heterogeneous architectures consisting of processors, data parallel and spatial compute units. The objective is to develop feasible models for the above DoD missions and accelerate them to support on the fly decision support. The proposed models are envisioned to be computationally efficient and capable of addressing challenging problems in decision science. For example, by integrating Graph Neural Networks (GNNs) with other deep neural networks and visualization techniques, we can provide interpretive AI that can help humans to make informed decisions. We have already developed state of the art models using GNNs for SAR ATR applications. These models have low complexity and are well suited for in flight processing of radar signals. In addition, the models have been trained to be robust against several classes of adversarial attacks.