Application of synthetically trained three-dimensional U-Net to the detection of moving subpixel objects
Katherine E. Eismann, Kenneth Mark Hopkinson, Bryan J. Steward, Shannon R. Young · Optical Engineering · 2025
The state-of-the-art advancements in passive sensors have coupled with machine learning for many applications. We apply machine learning to search for moving objects in infrared search and track applications. At long ranges, these objects appear unresolved, and conventional algorithms accumulate false alarms. As the background levels increase, this problem becomes very challenging. Machine learning offers the potential for better performance. We build upon prior research and thoroughly evaluate a three-dimensional U-Net algorithm against a conventional method in a proof-of-concept case, as well as more complex scenarios, including noise, stationary background clutter, and moving cloud clutter with dynamically moving objects and Earth backgrounds. The Air Force Institute of Technology Sensor and Scene Emulation Tool simulates the data as if collected from a downward-looking sensor while also providing labeled truth. A large, diverse, and realistic training dataset with associated truth ensures that the results represent the actual performance on real data. The results indicate comparable U-Net detection performance relative to statistical methods at between one and six times lower signal-to-noise or signal-to-clutter-plus-noise levels. We will test the capabilities of U-Net to take electro-optical sensor data and produce true object locations with high performance.