Open-set classification and network calibration for multisensor fusion in the transfer learning domain
Sudarshan Chakravarthy, Kameron Grubaugh, Christopher Ebersole, Mark Ashby, Edmund G. Zelnio · 2025
Advancements in the field of machine learning allow deep learning classifiers to perform exceptionally well in closed set problems; however, they lack robustness in open set problems. Moreover deep learning classifiers generally lack confidence estimates that accurately reflect their performance. These deep learning classifiers require large datasets, which are difficult and costly for the United States Air Force (USAF) to collect for targets of interest. While synthetically generated Synthetic Aperture Radar (SAR) data can cover many operating conditions (OCs) at a low cost, it still lacks statistical similarity when compared to measured SAR data. To address this challenge, fusion between multiple modalities can be employed to enhance the robustness and accuracy of these classifiers. Experimental results show the ability to perform fusion between passive Radio Frequency (RF) and SAR data in difficult open-set scenarios while still achieving high accuracy.