Transfer Learning and Deep Transfer Learning for Military Applications

Aided Target Recognition (AiTR) algorithms augment human decision-making, but require a large database of labeled targets for training. Applied to new domains, these algorithms fail to transfer knowledge and require substantial retraining. Areté and University of Alabama Huntsville (UAH) propose the development of the Domain Extracted Feature Transfer (DEFT) Network to transfer knowledge from a source to target domain. The ultimate DEFT product will be a cross-domain AiTR operating on a ground station. As multiple sensors provide new real-time data, DEFT will transfer knowledge to detect the same object across sensor modalities. During the course of Phase I, Areté will conduct a feasibility study to develop, train and evaluate DEFT on synthetic and measured datasets using an EO source and SAR target domain. During Phase II, Areté will further develop and demonstrate a prototype of the DEFT network and mature the algorithm for real world data (sparse, noisy, or imbalanced).