Transfer Learning-Based Fully-Polarimetric Radar Image Classification with a Rejection Option

In this paper, convolutional neural networks based on transfer learning are employed for the classification of fully polarimetric radar images with a rejection option. In a conventional supervised classification problem, the network has to choose from one of the known classes. However, a classifier, in a real-world scenario, may deal with an open set recognition problem in which the target under test is not included in the classifier training set. The capability of a classifier to discriminate between known and unknown targets may enable the classifier to self-learn from the experience as well as to enrich the system memory, which may in turn open the door to cognitive classifiers. This paper aims at testing the capability of transfer learning approaches, using VGG16 and AlexNet, to recognize unknown target classes of ISAR images. A threshold-based scheme has been applied to the softmax scores to discriminate among known and unknown targets. Data augmentation techniques such as translation and Gaussian noise addition are also employed in order to expand the diversity of the training dataset.