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

Physical Sciences Inc. (PSI) proposes the continued development of the Detection in IR with Incremental Low-shot Learning (DIRILL) algorithm suite, which integrates state-of-the-art machine learning object detection and classification capabilities with thermal IR sensors using low-shot training techniques. DIRILL will ultimately interface with existing Army platforms to support Aided Target Recognition (AiTR) on deployed combat vehicles with modern thermal infrared (IR) sensors. In the successful Phase I program, PSI demonstrated the feasibility of training machine learning models for novel object detection and identification with >95% probability of detection (PD) and >90% probability of identification (PID) at 0.01 false alarms per image (FAR) using as few as 30 labeled training images per class on government-provided ground-to-ground and air-to-ground datasets containing military targets. The DIRILL system uses self-supervised learning and feature reweighting to achieve this performance with IR sensors where limited labeled training datasets are available. The self-supervised learning on unlabeled imagery enables direct training on IR data with no requirement to conform to image formats of existing labeled visible imagery datasets, improving performance by leveraging the full dynamic range and separate phenomenology of IR sensors. The algorithms enable the rapid (<25 minutes of training time) detection and identification of an easily-updated set of targets appropriate for specific missions and operation environments using minimal data (10s – 100 labelled images). The final DIRILL capability will allow combat personnel to rapidly identify both friendly forces and uncooperative threats using AiTR in new and complex scenarios. DIRILL maintains compatibility with PSI’s Low-shot Object Tracking and Targeting (LOTT) software modules being developed under a separate Phase II Army SBIR program (Contract #W909MY-21-C-0010). DIRILL is intended to replace the current object detector algorithm while leveraging the downstream targeting and tracking capabilities of LOTT.