講演抄録/キーワード |
講演名 |
2018-10-12 13:30
Automatic Target Recognition based on Generative Adversarial Networks for Synthetic Aperture Radar Images Yang-Lang Chang・Bo-Yao Chen・○Chih-Yuan Chu・Sina Hadipour(NTUT)・Hirokazu Kobayashi(OIT) SANE2018-51 |
抄録 |
(和) |
Synthetic Aperture Radar (SAR) is an all day and all weather condition imaging technique which is widely used in national defense, remote sensing, disaster prevention, interferometry and forest and urban footprint mapping. Recently, convolutional neural networks have been used for automatic target recognition (SAR-ATR) and classification. The drawback, however, is the difficulty obtaining sufficient and reliable data in order to train a high accuracy classifier for automatic target recognition. As the number of training samples is reduced, the SAR-ATR accuracy rate decreases rapidly. Our study proposes a deep learning model based on Generative Adversarial Network (GAN) to overcome the problem of insufficient training samples and improve the performance of target classification. GAN is composed of two networks: A Generator network and a Discriminator network. The generator network produces SAR images from a series of random numbers. The discriminator network is a classifier which is trained using supervised learning to classify real and fake SAR images. The Generator and the Discriminator compete with each other in the training process in order to learn robust and reliable target features in SAR images. However, traditional GAN cannot be used to solve the classification problems in SAR-ATR. Our network is a variant of GAN called Auxiliary Classifier GAN (AC-GAN). The structure of AC-GAN allows separating large datasets into subsets by class and training a generator and discriminator for each subset. In this experiment, the SAR images in Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset were used to train the network. Using all the images in the dataset for training resulted in a classification accuracy of 98%. When only less than one-fifth of the images were used, AC-GAN reached an accuracy of 90%. This is a considerable increase in accuracy in comparison with traditional CNNs where for the same number of training samples, the accuracy rapidly decreased to 80%. |
(英) |
Synthetic Aperture Radar (SAR) is an all day and all weather condition imaging technique which is widely used in national defense, remote sensing, disaster prevention, interferometry and forest and urban footprint mapping. Recently, convolutional neural networks have been used for automatic target recognition (SAR-ATR) and classification. The drawback, however, is the difficulty obtaining sufficient and reliable data in order to train a high accuracy classifier for automatic target recognition. As the number of training samples is reduced, the SAR-ATR accuracy rate decreases rapidly. Our study proposes a deep learning model based on Generative Adversarial Network (GAN) to overcome the problem of insufficient training samples and improve the performance of target classification. GAN is composed of two networks: A Generator network and a Discriminator network. The generator network produces SAR images from a series of random numbers. The discriminator network is a classifier which is trained using supervised learning to classify real and fake SAR images. The Generator and the Discriminator compete with each other in the training process in order to learn robust and reliable target features in SAR images. However, traditional GAN cannot be used to solve the classification problems in SAR-ATR. Our network is a variant of GAN called Auxiliary Classifier GAN (AC-GAN). The structure of AC-GAN allows separating large datasets into subsets by class and training a generator and discriminator for each subset. In this experiment, the SAR images in Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset were used to train the network. Using all the images in the dataset for training resulted in a classification accuracy of 98%. When only less than one-fifth of the images were used, AC-GAN reached an accuracy of 90%. This is a considerable increase in accuracy in comparison with traditional CNNs where for the same number of training samples, the accuracy rapidly decreased to 80%. |
キーワード |
(和) |
synthetic aperture radar / automatic target recognition / generative adversarial networks (GAN) / auxiliary classifiers GAN / / / / |
(英) |
synthetic aperture radar / automatic target recognition / generative adversarial networks (GAN) / auxiliary classifiers GAN / / / / |
文献情報 |
信学技報, vol. 118, no. 239, SANE2018-51, pp. 41-44, 2018年10月. |
資料番号 |
SANE2018-51 |
発行日 |
2018-10-05 (SANE) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
SANE2018-51 |
|