(2) Deep Learning-based UAS Multimodal Imagery Perception

To address the Navy’s need for the development of cutting-edge artificial intelligence (AI)/machine learning (ML) techniques for accurate unmanned aircraft system (UAS) image recognition, Physical Optics Corporation (POC) proposes to develop a new Deep Learning-based UAS Multimodal Imagery Perception (DUMIP) software suite. It is based on a new system design that utilizes convolutional neural networks (CNNs) tailored for resource-constrained environments that require low size, weight, and power processing hardware. The innovation in a multimodal fusion neural network to fuse visible and infrared imagery capable of handling modality drop, transfer learning with customized image quality, and a novel scene reasoner capable of improving detection accuracy will enable DUMIP to improve the classification of the environment and enhance target recognition for UASs with very high accuracy. This system offers advanced image understanding techniques for handling low-quality images and multimodal imagery (sensor fusing), directly addressing the Navy’s requirement for enhanced image understanding with multimodal inputs. In Phase I, POC will demonstrate the feasibility of DUMIP by training and testing DUMIP networks followed by a technology demonstration that will reach technology readiness level (TRL)-3. In Phase II, POC plans to develop, demonstrate functionality and deliver DUMIP prototype system for testing and evaluation, reaching TRL-5.