本研究將電磁訊號模擬結果結合深度學習應用於軍事用途，期望透過雷達蒐集到的雷達散射截面積(Radar Cross Section, RCS)數據快速辨識接近的不明船型目標物。因為在作戰期間不太會有將船型目標物的各角度RCS數據都量測出來的機會，加上實際量測環境條件會比模擬情況還嚴苛，本研究討論不同的RCS測試數據集加入不同大小的雜訊後對神經網路辨識的影響，以及透過生成對抗網路(Generative Adversarial Network, GAN)來進行數據增強(Data Augmentation)觀察是否能改善原有的神經網路辨識。
本研究的第一部分先討論如何透過CST MIRCROWAVE STUDI(CST MWS)來蒐集模擬船型目標物的RCS數據集，數據集根據不同的蒐集方法分為角度分集法與頻率分集法，接著分別對測試數據集加入不同大小的高斯亂數模擬實際量測環境中的雜訊干擾並觀察對神經網路辨識的影響，本研究室使用卷積神經網路(Convolutional Neural Network ,CNN)來進行船型目標物的RCS數據辨識。
本研究的第二部分使用生成對抗網路作為數據增強的手段，因為RCS數據取得困難且費時費力，而透過用生成對抗網路來進行RCS數據生成可以解決深度學習訓練數據集不足導致辨識效果不理想的窘境。訓練數據集加入生成的RCS數據進行訓練後，同樣的分別對測試數據集加入不同大小的高斯亂數模擬實際量測環境中的雜訊干擾並觀察加入生成對抗網路生成的RCS數據是否能改善辨識的效果。從辨識結果來看，隨著雜訊的增強，在加入雜訊的測試數據集的辨識準確率可能降低至41%，而加入生成的RCS數據後在測試數據受雜訊干擾小時辨識準確率比沒有加入生成RCS數據最多可提升1.87%，但當測試數據受雜訊干擾大到一個程度時就失去辨識準確率提升的效果。
關鍵字：雷達目標辨識 ; 雷達散射截面積 ；卷積神經網路 ; 生成對抗網路

The purpose of this study is to improve the performance of Convolutional Neural Network(CNN) models for electromagnetic signals recognition of ship targets using Generative Adversarial Network(GAN) based Data Augmentation. Radar cross section(RCS) of a target is the equivalent area seen by a radar and is also called electromagnetic signature of the object. RCS datasets include several different diversities depending on collection data method. In this study, angle dataset and frequency dataset were adopted. CST MICROWAVE STUDIO (CST MWS), a high-performance 3D electromagnetic analysis software, was used to build RCS datasets by simulating RCS of 5 different ship models. The effects of ocean wave were also considered in the simulation. To simulate the noise that often occurs under realistic conditions, Random numbers drawn from Gaussian Distributions with different standard deviations were added to the test dataset.
RCS data of ships are insufficient due to the difficulty of measuring RCS of ships at every angle. GAN-based Data Augmentation is regarded as a solution for the problem of limited data. The RCS data generated by GAN were added to training dataset for increasing the size of training dataset. The result indicates that the recognition accuracy increases by up to 1.87 percent after using GAN-based Data Augmentation, and the improvement on the frequency dataset is better than angle dataset.
Key words: Radar Target Recognition, Radar Cross Section, Convolutional Neural Network, Generative Adversarial Network

[1] X. Jiang, A. Hadid, Y. Pang, E. Granger, and X. Feng, Deep Learning in Object Detection and Recognition, Springer, 2019
[2] I. Harre, RCS in Radar Range Calculations for Maritime Targets, Accessed on: Jun. 10, 2021. [Online]. Available: http://www.mar-it.de/Radar/RCS/RCS_xx.pdf
[3] I. J. Goodfellow et al., “Generative Adversarial Nets,” In Advances in Neural Information Processing Systems, pp. 2672–2680, 2014
[4] A. Krizhevsky, I. Sutskever and G. E. Hinton, “Imagenet classification with deep convolutional neural networks," Proc. Neural Inf. Process, Syst. 25, pp. 1106-1114, 2012
[5] P. Blacksmith, Jr. R. E. Hiatt, and R. B. Mack, “Introduction to radar cross-section measurements,” Proc. IEEE, vol. 53, pp. 901-920, 1965
[6] M.B. Perotoni, and S.E. Barbin, “A Study on RCS of Missile Models using the Method of Moments,” IEEE MTT-S International Microwave & Optoelectronics Conference, pp. 492-495, 2007
[7] D. Song et al., “Electromagnetic Analysis in the Ship RCS Plastics Design,” Chinese Journal of Ship Research, Vol. 4, No. 3, pp.52-55, 2009
[8] G. Zhang et al., ” RCS model of ship based on FEKO,” International Conference on Computational Problem-Solving (ICCP), pp. 419-422, 2013
[9] Ç. Durlu, and H. T. Hayvaci, “The analysis of RCS of winged almond model with ammunition and without ammunition,” In International Artificial Intelligence and Data Processing Symposium (IDAP), 2017
[10] J. Shi, R. Chen, Y. Zuo, and A. Jiao, “Estimation of the Ship RCS on the Dynamic Sea Surface,” Photonics & Electromagnetics Research Symposium - Fall (PIERS - Fall), pp. 1755-1759, 2019
[11] S. Hao et al., “Research on the shape stealth design of infantry Fighting vehicle,” IEEE International Conference on Intelligent Computing and Signal Processing (ICSP), pp.1393-1396, 2021
[12] N. Ødegaard et al., “Classification of Ships Using Real and Simulated Data in a Convolutional Neural Network,” In IEEE Radar Conference, 2016
[13] S. Xing, and S. Zhang, “Ship Model Recognition Based on Convolutional Neural Networks,” IEEE International Conference on Mechatronics and Automation, pp. 144-148, 2018
[14] C. Shorten, and T.M. Khoshgoftaar, “A survey on Image Data Augmentation for Deep Learning,” Journal of Big Data, 6(1):60, 2019
[15] A. Radford, L. Metz, and S. Chintala, “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks,” In Proceedings of the International Conference on Learning Representations (ICLR), 2016
[16] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein GAN,” arXiv preprint arXiv:1701.07875, 2017
[17] J. Ma et al., “DDcGAN: A Dual-Discriminator Conditional Generative Adversarial Network for Multi-Resolution Image Fusion,” In IEEE Transactions On Image Processing, vol. 29, 2020
[18] A. Hammoudeh, and M. Nuaimi, “Scattering cross-section characterisation of vegetation media at microwave frequencies,” IET Journals & Magazines, Electronics Letters, vol. 30, no. 11, pp. 903-904, 1994
[19] H. Hirayama, H. Hayashi, N. Kikuma, and K. Sakakibara, “Estimation of poynting vector and wavenumber vector from near-magnetic-field measurement,” In IEEE Conferences, 12th International Symposium on Antenna Technology and Applied Electromagnetics and Canadian Radio Sciences Conference, 2006
[20] A. Pathak, Z. Zaheeruddin, and M. K. Tiwari, “Minimizing the energy hole problem in wireless sensor networks by normal distribution of nodes and relaying range regulation,” IEEE Fourth International Conference on Computational Intelligence and Communication Networks, pp. 154-157, 2012
[21] Y. Grishin, and K. Konopko, “Faults diagnosis of stochastic dynamic systems based on neural network probability density function estimation,” Proceedings of the 12th IEEE Mediterranean Electrotechnical Conference, pp. 335-338, 2004
[22] K. Horiuchi, and K. Kameyama, “Parameter density inheritance using kernel density estimation for efficient CNN learning,” IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), pp. 308-313, 2018