Airborne Video Inspection for Automatic Targeting with Ontology Reasoning (AVIATOR)

Improved cameras and unmanned air vehicles (UAVs) have led to an explosion in the amount of airborne imagery and video collected by Air Force assets, but there are not enough trained personnel to analyze this imagery in real time. Thus, targets of opportunity and threats go undetected until the chance to act on them has passed. Automatic target detection could alleviate the burden on analysts and reduce the time required to detect and prosecute targets of opportunity. Deep learning techniques have revolutionized computer vision and established state of the art results in detection and classification. However, deep neural networks (DNNs) are black boxes of linear algebra that make inscrutable decisions. To bring the power of deep learning to bear on airborne image exploitation in a way that humans can understand and trust, we propose Airborne Video Inspection for Automatic Targeting with Ontology Reasoning (AVIATOR). AVIATOR uses deep learning to maximize airborne detection accuracy, but critically it uses a “bolt-on� explanation system that analyzes a DNN to provide end users with intuitive explanations for its decisions via a web-based graphical interface. These accurate human-understandable detections will gain the trust of users and support after action reviews to facilitate adoption.