Synthetic aperture radar automatic target recognition (SAR ATR) technology is widely utilized in military target detection. However, the difficulty in obtaining labeled SAR samples restricts the utilization of existing recognition techniques. A SAR target recognition method, which combines a wasserstein GAN with gradient penalty (WGAN-GP) with a pre-trained model, is proposed. After augmentation of small training datasets using WGAN-GP, the data is then fed into a convolutional neural network (CNN) model pretrained on the large-scale remote sensing image scene classification (RESISC) dataset for training, ultimately yielding SAR target recognition results. The algorithm’s capabilities are evaluated using the moving and stationary target acquisition and recognition (MSTAR) dataset. Experimental results indicate WGAN-GP, when utilized, outperforms other generative adversarial networks in SAR sample enhancement. Furthermore, the selection of the RESISC45 dataset is found to effectively enhance the classifier’s pretraining ability. Compared to existing research findings, the approach exhibits advantages in improving SAR target recognition accuracy and CNN model convergence speed.
由于难以提取SAR图像的底层特征和GAN本身存在的缺陷,以上方法所使用的GAN在训练过程中都存在不稳定的风险。考虑到带梯度惩罚的生成对抗网络(Wasserstein GAN with Gradient Penalty, WGAN-GP)[14]从原理上给出了应对方案,能很好地解决此类问题,提出一种基于WGAN-GP和大规模数据集预训练的SAR-ATR方法。首先利用WGAN-GP对不同种类的小样本训练数据集进行分别扩充,得到新的带标签扩充数据集;然后将合并的数据集导入具有大规模数据集预训练权重的CNN模型中进行训练;最终得到SAR目标识别结果。利用MSTAR数据集[15]上进行检测,在训练样本有限的情况下,基于WGAN-GP和大规模数据集预训练方法的识别精度和模型收敛速度得到了显著提高,并有着良好的泛化效果。
DUDGEOND, LACOSSR. An overview of automatic target recognition[J]. The Lincoln Laboratory Journal,1993,6(1):3-10.
[3]
PARKJ I, KIMK T. Modified polar mapping classifier for SAR automatic target recognition[J].IEEE Transactions on Aerospace and Electronic Systems, 2014,50(2):1092-1107.
[4]
HUMMELR. Model-based ATR using synthetic aperture radar[C]∥Record of the IEEE 2000 International Radar Conference. Alexandria,Egypt: IEEE, 2000:856-861.
[5]
KNAPSKOGA O.Classification of ships in TerraSAR-X images based on 3D models and silhouette matching[C]∥8th European Conference on Synthetic Aperture Radar.Aachen,Germany: VDE, 2010:1-4.
[6]
ZHAOQ, PRINCIPEJ C. Support vector machines for SAR automatic target recognition[J]. IEEE Transactions on Aerospace and Electronic Systems,2001,37(2):643-654.
[7]
SUNY J, LIUZ P, TODOROVICS,et al. Adaptive boosting for SAR automatic target recognition[J]. IEEE Transactions on Aerospace and Electronic Systems,2007,43(1):112-125.
[8]
KRIZHEVSKYA, SUTSKEVERI, HINTONG E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM,2017,60(6):84-90.
GUOJ Y, LEIB, DINGC B,et al. Synthetic aperture radar image synthesis by using generative adversarial Nets[J]. IEEE Geoscience and Remote Sensing Letters,2017,14(7):1111-1115.
[11]
GAOF, YANGY, WANGJ,et al. A deep convolutional generative adversarial networks(DCGANs)-based semi-supervised method for object recognition in synthetic aperture radar(SAR)images[J]. Remote Sensing,2018,10(6):846.
[12]
ZHENGC, JIANGX, LIUX Z. Multi-discriminator generative adversarial network for semi-supervised SAR target recognition[C]∥2019 IEEE Radar Conference.Boston,USA: IEEE,2019. DOI:10.1109/RADAR.2019.8835801 .
[13]
LIUX Y, HUANGY L, WANGC W,et al. Semi-supervised SAR ATR via conditional generative adversarial network with Multi-Discriminator[C]∥2021 IEEE International Geoscience and Remote Sensing Symposium. Brussels国家:IEEE,2021:2361-2364.
[14]
GULRAJANII, AHMEDF, ARJOVSKYM,et al. Improved training of wasserstein GANs[C]∥Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, California, USA:Curran Associates Inc.,2017:5769-5779.
[15]
KEYDELE R, LEES W, MOOREJ T. MSTAR extended operating conditions:a tutorial[C]∥Algorithms for Synthetic Aperture Radar Imagery III.Orlando,USA:SPIE,2019. DOI:10.1117/12.242059 .
[16]
CHENGG, HANJ W, LUX Q.Remote sensing image scene classification: benchmark and state of the art[J]. Proceedings of the IEEE, 2017, 105(10):1865-1883.
RENS Q, HEK M, GIRSHICKR,et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017,39(6):1137-1149.
[19]
AHARONM, ELADM, BRUCKSTEINA. K-SVD: an algorithm for designing overcomplete dictionaries for sparse representation[J]. IEEE Transactions on Signal Processing, 2006,54(11):4311-4322.