Automatic Target Recognition of Synthetic Aperture Radar Images Using Deep Learning Techniques

Abstract

Maha Al Mufti. “Automatic Target Recognition of Synthetic Aperture Radar Images Using Deep Learning Techniques.' MSc. Thesis by Research in Engineering, Department of Electrical and Computer Engineering, Khalifa University of Science and Technology, United Arab Emirates, July 2018. Synthetic Aperture Radars (SARs) are spaceborne or airborne imaging radar systems. SARs have the unique characteristics of creating high resolution images in all-weather condition and not depending on the availability of sunlight. These systems have many different applications including military surveillance applications or civilian weather applications. SARs detect different objects and targets during their operation. Automatic Target Recognition (ATR) is the area of research that is concerned with automating the recognition process of these targets or objects. In this research, a deep learning based approach is proposed to tackle the SAR-ATR problem. The approach is based on a transfer learning scheme where pre-trained Convolutional Neural Networks (CNNs) are employed to transfer the previously gained knowledge and apply it to the SAR-ATR problem. Two transfer learning based schemes are experimented with, namely feature extraction and fine-tuning. Three CNNs, which are pre-trained on the ImageNet optical images dataset are, used in the course of this research namely AlexNet, VggNet and GoogleNet. The approach was validated on two SAR datasets, the Moving and Stationary Target Acquisition and Recognition (MSTAR) and the Bright Spark dataset. An analysis of the results shows that the knowledge gained by the pre-trained CNNs is transferable and applicable to the SAR-ATR problem achieving results that are comparable to the state-of-the-art. Indexing Terms: Deep learning, SAR, Transfer learning, ATR

Date of Award	Jul 2018
Original language	American English
Supervisor	Naoufel Werghi (Supervisor)