Sar image scene classification and out-of-library target detection with cross-domain active transfer learning

The majority of the existing deep-learning based SAR automatic target recognition (ATR) algorithms rely solely on the "appearance" of the SAR signatures for target classification, while ignoring the relationship between the objects of interest and their surroundings. In this work, we emphasize on enhancing the capability of SAR ATR algorithms in detecting and categorizing out-of-library (OOL) targets in open environment with context-based compositional learning. Rather than attempting to build a single do-it-all model, task-specific sub-models are chosen based on the natural selection mechanism, whose relationships are structured via logic flow based on context. To compensate for the SAR training data scarcity and the unequal distribution of classes, active learning, cross-domain transfer learning, and transductive learning are jointly exploited. Simulation results show that the proposed joint scene-target recognition framework could potentially solve the challenging problem of OOL target classification in complex mission scenarios.