ISAR Ship Classification Using Transfer Learning

In an airborne maritime radar, inverse synthetic aperture radar (ISAR) is used to image and classify noncooperative targets. Traditional classification approaches rely on geometric features extracted from images of known targets to form a training dataset that is later used to classify observed targets. In recent years, deep learning-based techniques have been applied to a number of radar problems with demonstrated improvements over conventional processing schemes. The application to ISAR image classification is difficult due to the availability of small training datasets and the inability to classify vessels of an unknown class. In this work, we propose a transfer learning approach to address the small data problem, while the unknown class issue is addressed with the use of an output layer known as OpenMax. Using an ISAR dataset of small vessels, the new classification results are compared with a traditional classification approach and a simple three-layer Convolutional Neural Network (CNN). We have observed that the use of OpenMax to classify the images of vessels from an unknown class has improved classification performance significantly.

Automated summary

Inverse synthetic aperture radar (ISAR) is increasingly used in airborne maritime radar for noncooperative target imaging and classification. Traditional classification methods are limited by their reliance on geometric features extracted from images of known targets, which hampers their ability to classify unknown vessels. To address this challenge, this study proposes a transfer learning approach combined with an output layer called OpenMax. By comparing the new classification results with traditional methods and a three-layer Convolutional Neural Network (CNN) using a dataset of small vessels, it is observed that the use of OpenMax significantly improves classification performance for vessels from unknown classes.