Binarized Deep Fusion Classification

To address the Navys need for a radar and electro-optical/infrared (EO/IR) fusion system running on space, weight and power (SWaP) constrained platforms for ship classification and identification, Physical Optics Corporation (POC) proposes to develop a new Binarized Deep Fusion Classification (BDFC) technology. This proposed BDFC solution is based on decision-level and feature-level data fusion of multiple sensory modalities using artificial neural network (ANN) classifiers. Imaging modalities are classified by the state-of-the-art convolutional neural networks (CNNs). A policy network fuses the output of the individual classifiers and metadata to make the final decision. BDFC utilizes innovative ANN binarization to facilitate its real-time implementation on mobile computing devices and field-programmable gate arrays (FPGAs). As a result, BDFC offers rapid sensor fusion on low-SWaP platforms for high-confidence target classification, which directly addresses the PMA-299 requirements. In Phase I, POC will reach TRL-3 by training and testing the model on synthetic datasets to demonstrate the enhanced classification accuracy by fusing multi-sensor modalities at a 5 Hz solution rate on a laptop. Transition to low-SWaP platforms will be planned. In Phase II, POC will reach TRL-6 by implementing and optimizing the algorithm for the chosen computing platform and testing it on real-world datasets.