Machine Learning Approaches for Data Limited SAR Object Recognition

In this lecture, we present machine learning approaches to classify data limited SAR imagery. Recently, active learning algorithms have been applied to video imagery labeling and classification. Active learning enables labeling a small subset of imagery by hand/human, learning features of an object from the labeled imagery, and then applying learned features to automatically label/classify unlabeled/new imagery.

The goal here is that instead of using lots of data (~70%) for training, can we use only a fraction (10-15%) of the data for training and thus reduce training time and learn new objects much faster.  Similar algorithms named “few-shot learning (FSL)” or “low-shot learning (LSL) also enable learning objects from a few samples of training data. We will present active learning and few-shot learning algorithms to classify SAR imagery.