Data Density Neural Networks

As AI proliferates into everything from satellite imagery, soldier worn systems, drones, and many others, carefully designed neural networks tuned for specific targets are being evaluated in true field conditions.  This proposal, Data Density Neural Networks (DDNN) is an effort to resolve a common failure mode of AI technologies in real field data: high contrast imagery. High contrast imagery affects AI as much as it does people – a good example is bright flashlights in true darkness, or camera systems adjusting to direct sunlight outdoors. This problem has two broad roots: training data for an AI object detector or classifier is often not collected with real field conditions in mind, and camera systems optimize auto gain and auto exposure to create a picture more suited for a person, not for data density and AI. DDNN is a project to develop a data density model inherent to the neural network structure that provides existing camera/AI systems enhanced performance in very dark and very light areas. Those ares where data density issues are inherent to camera behavior in high contrast environments. For this Phase I AFWERX project, a technical development producing a side-by-side comparison of a preliminary DDNN method with conventional AI techniques will be produced, along with a USAF customer discovery task with our DoD transition partner. The three month project will deliver the side-by-side video analysis, a preliminary and final report, and completed forms SF298 and DD822.