Incremental Learning in Synthetic Aperture Radar Images Using Openmax Algorithm

Abstract

In real-time real-world scenarios, an automatic target recognition (ATR) system may encounter new samples from unseen classes continually. Retraining a neural network by using the new and all the previous samples, whenever new data is received, imposes a considerable computational cost. Instead, incremental learning aims at learning new knowledge while preserving previous knowledge with an emphasis on computational time and storage resources. In this paper, we employ the Openmax method, which has been initially introduced for open set recognition in optical images, to assist a convolutional neural network (CNN) in incremental learning scenarios with SAR images. The new set for fine-tuning the network is constituted of the unknown samples recognized by the Openmax method together with exemplars from the old classes. Our real data analysis to validate the proposed method is performed on radar images of man-made targets from the well-known Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset.