Classification of ISAR Ship Imagery Using Transfer Learning
Luke Rosenberg, Weiliang Zhao, Anthony Heng, Len Hamey, Si Tran Nguyen Nguyen, Mehmet A. Orgun
Abstract
Inverse synthetic aperture radar (ISAR) is a common radar imaging technique used to characterize and classify non-cooperative targets. Traditional classification approaches use geometric features extracted from the images of known targets to form a training dataset that is later used to classify an unknown target. While these approaches work reasonably well, deep learning-based techniques have demonstrated significant improvements over conventional processing schemes in many areas of radar. However, the application of ISAR image classification is difficult when there are only small training datasets available. In this article, we address the small dataset problem by utilizing transfer learning. Different approaches are considered that can take advantage of the ship aspect angle to improve the overall stability and improve the final classification result. The new classification results are then compared with a traditional classification approach and a simple three-layer convolutional neural network. In addition, to better understand how the neural networks are working, saliency maps are used to visualize the trained network.