Detection in IR with Incremental Low-shot Learning (DIRILL)

Physical Sciences Inc. (PSI) proposes to further develop, mature and transition generalizable and adaptable Aided Target Recognition (AiTR) algorithmsĀfor Army Launched Effects (LE) systems. Under the previous successful Phase II effort, we developed and demonstrated Detection in Infrared (IR) with Incremental Low-shot Learning (DIRILL) algorithms that implement semi-supervised training techniques to minimize labeled data requirements and support rapid adaptation for new targets and new multispectral IR and visible sensors. DIRILL demonstrated 94% PD at <1 false detection per 10 images analyzed and 91% PID for civilian/military targets in operationally relevant untrained scenes. When finetuned on 25 ¢ 100 images of three new targets in 15 ¢ 30 minutes on an embedded GPU platform, DIRILL achieved 95% PD and 79% PID for these previously untrained targets in new scenes. PlannedĀalgorithm improvements include automation of the model adaptation process to minimize operator involvement, mitigations for degraded visual information, optimal model scaling to enable highest performance at various system constraints, and integration with higher level autonomy systems. PSI will work closely with Army RTI to develop a Modular Open Systems Architecture (MOSA) compliant algorithm payload compatible with current LE platforms, and demonstrateĀthe capability at key RTI and Army PEO AviationĀdemonstration events.