Integrated GANs: Semi-Supervised SAR Target Recognition

Fei Gao,Amir Hussain,Jinping Sun,Huiyu Zhou,Qiuyang Liu

doi:10.1109/access.2019.2935167

Abstract

With the advantage of working in all weathers and all days, synthetic aperture radar (SAR) imaging systems have a great application value. As an efficient image generation and recognition model, generative adversarial networks (GANs) have been applied to SAR image analysis and achieved promising performance. However, the cost of labeling a large number of SAR images limits the performance of the developed approaches and aggravates the mode collapsing problem. This paper presents a novel approach namely Integrated GANs (I-GAN), which consists of a conditional GANs, an unconditional GANs and a classifier, to achieve semi-supervised generation and recognition simultaneously. The unconditional GANs assist the conditional GANs to increase the diversity of the generated images. A co-training method for the conditional GANs and the classifier is proposed to enrich the training samples. Since our model is capable of representing training images with rich characteristics, the classifier can achieve better recognition accuracy. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset proves that our method achieves better results in accuracy when labeled samples are insufficient, compared against other state-of-the-art techniques.

Highlights

Synthetic aperture radar (SAR) imaging systems adopt coherent imaging principles, which can effectively penetrate clouds and collect rich target information
Our main contributions are summarized as follows: (1) We investigate the problems of the existing models based on generative adversarial networks (GANs) for SAR image recognition, and propose a novel GANs model for semi-supervised learning
The fake SAR images generated by cG in the training process are shown in Fig. 5, where the numbers of the labeled SAR images are 600 and 100 respectively

Summary

Introduction

Synthetic aperture radar (SAR) imaging systems adopt coherent imaging principles, which can effectively penetrate clouds and collect rich target information. High resolution SAR remote sensing systems, such as Terra SAR-X [1] and COSMO-SkyMed [2], have been widely used in aerial surveying and space reconnaissance. SAR image recognition technology has become an active research area in SAR image remote sensing [3]. Traditional recognition methods for SAR images includes Template Matching [4], Support Vector Machine (SVM) [5], and Adaptive Boosting (AdaBoost) [6]. These traditional methods are largely dependent on hand-crafted features. When image data is large and complex, The associate editor coordinating the review of this article and approving it for publication was Gerardo Di Martino

Methods

Results

Conclusion