Incremental learning using manifold feature space for synthetic aperture radar target recognition

Abstract

While deep neural network technology brings high recognition accuracy to the field of synthetic aperture radar automatic target recognition, it also produces the problem of catastrophic forgetting. Currently, how to extract features for distinguishing new and old classes has become the main bottleneck for incremental learning performance improvement. In this paper, we propose a new incremental learning method to better distinguish between new and old classes. We use the trained neural network to extract the features of the old samples and utilize the k -means to select representative old samples in the feature space, and then train the new model with distillation loss. Through the experiments on the MSTAR dataset, our method has better incremental learning performance on SAR images under the same training time.