Blending Classical Model-Based Target Classification and Identification Approaches with Data-Driven Artificial Intelligence

Toyon Research Corp. and the University of California propose to develop innovative algorithms to perform automatic target recognition (ATR), localization, and classification of maritime and land targets in EO/IR, LiDAR, and SAR imagery. The proposed algorithms are based on recent developments made at the University of California, which outline a strong mathematical framework for naturally blending classical model-based target recognition algorithms with modern data-driven machine-learning approaches, such as those utilizing deep learning. Specifically, this proposal presents a mathematical formalism for incorporating heuristics and feature-based detection algorithms directly into a deep neural networks (DNNs), which can be further trained and optimized for a particular ATR task specified by example or model-data. The resulting system provides a mechanism for Navy operators and DoD algorithm designers to leverage existing model-based ATR algorithms, with a known threshold of baseline-performance, to systematically design deeper and more-robust neural-network-based algorithms that can be naturally optimized and augmented using modern machine-learning approaches. Furthermore, while the proposed algorithmic framework is largely agnostic to the particular imaging-modality, we demonstrate how it can be specifically tuned to discover features that are unique to particular modalities, in either the spatial or the spatial-frequency domain, mitigating uncertainties that arise in various environmental or