Vladimir V. Razevig, Aleksandr I. Ivashov, Aleksandr S. Bugaev

 Bauman Moscow State Technical University (Moscow, Russia), Moscow Institute of Physics and Technology (State University) (Moscow, Russia)

 vrazevig@rslab.ru

Subsurface radars, emitting radio waves and registering the reflected signal, make it possible to obtain images of objects in an optically opaque medium, or the internal structure of objects that are opaque in the optical wavelength range. The acquisition of images (called microwave images) is based on the processing of microwave holograms (a microwave hologram is an ordered set of samples of the complex amplitude of the reflected signal recorded at a certain aperture).

The lateral resolution of subsurface radars using classical microwave hologram reconstruction methods such as back projection method and back propagation method is limited by a diffraction limit of approximately half the wavelength of the signal being used. Therefore, the efforts of many researchers are directed to the development of methods and algorithms to overcome the diffraction limit and obtain superresolution. 

The use of artificial intelligence (artificial neural networks), which is one of the leading research areas of scientists today, can help in solving this problem. 

In this paper, a neural network method for reconstructing microwave holograms is investigated. In the training sample, the input data were microwave holograms of various objects, which were modeled using sets of densely spaced point scatterers. Arrays of true scatterer positions were used as output data.

It was shown that the neural network is indeed capable of learning how to reconstruct microwave holograms: it demonstrated correct reconstruction results even for objects that were not present in the training sample.

The results of the neural network were compared with the results of reconstruction by the traditional back projection method. The neural network is capable of obtaining microwave images with better resolution than back projection method, and this resolution can be called superresolution, since it exceeds the diffraction limit.

The disadvantage of the developed neural network is a noticeable level of background artifacts (compared to back projection method), which can be reduced by using in the output layer instead of the identical activation function, for example, a quadratic,  exponential, or threshold function.

Razevig V.V., Ivashov A.I., Bugaev А.S. Neural network method for microwave holograms reconstruction. Achievements of modern radioelectronics. 2020. V. 74. № 12. P. 5–16. DOI: 10.18127/j20700784-202012-01. [in Russian]

1. Ivashov S.I., Razevig V.V., Vasiliev I.A., Zhuravlev A.V., Bechtel T.D., Capineri L. Holographic Subsurface Radar of RASCAN Type:  Development and Applications. IEEE Journal of Selected Topics in Earth Observations and Remote Sensing. December 2011. V. 4. № 4. P. 763–778.
2. Kuriksha A.A. Algoritm obratnoy proektsii v zadachakh vosstanovleniya prostranstvennogo raspredeleniya istochnikov voln. Radiotekhnika i elektronika. 2002. T. 47. № 12. S. 1484–1489. [in Russian]
3. Sheen D.M., McMakin D.L., Hall T.E. Three-dimensional millimeter-wave imaging for concealed weapon detection. IEEE Transactions on Microwave Theory and Techniques. 2001. V. 49. № 9. P. 1581–1592.
4. Daniels D.J. Ground Penetrating Radar, 2nd ed. (IEE, London, 2004).
5. Neyronnaya set'. Bol'shaya rossiyskaya entsiklopediya: [v 35 t.]. gl. red. Yu.S. Osipov. M.: Bol'shaya rossiyskaya entsiklopediya. 2004– 2017. [in Russian]
6. Shah P., Moghaddam M. Super resolution for microwave imaging: A deep learning approach. 2017 IEEE International Symposium on Antennas and Propagation & USNC/URSI National Radio Science Meeting, San Diego, CA. 2017. P. 849–850.
7. Zhang Y., Xiao Z., Wu L., Lu X., Wang Y. Deep learning for subsurface penetrating super-resolution imaging. 2017 10th UK-EuropeChina Workshop on Millimetre Waves and Terahertz Technologies (UCMMT), Liverpool, 2017. P. 1–4.
8. Panda S., Akhter Z., Akhtar M.J. Subsurface imaging of concrete structures using neural network approach. 2016 IEEE MTT-S International Microwave and RF Conference (IMaRC), New Delhi, 2016. P. 1–4.
9. Watabe K., Shimizu K., Yoneyama M., Mizuno K. Millimeter-wave active imaging using neural networks for signal processing. IEEE Transactions on Microwave Theory and Techniques. 2003. V. 51. № 5. P. 1512–1516.
10. Velik R. Discrete Fourier Transform Computation Using Neural Networks. 2008 International Conference on Computational Intel-ligence and Security, Suzhou. 2008. P. 120–123.
11. Rivenson Y., Zhang Y., Gunaydin H., Teng D., Ozcan A. Phase recovery and holographic image reconstruction using deep learning in neural networks. Light Sci. Appl. 2018. V. 7. Art. no. 17141.
12. Agostinelli F., Hoffman M., Sadowski P., Baldi P. Learning Activation Functions to Improve Deep Neural Networks, 21 Dec 2014. arΧiv:1412.6830 [cs.NE].
13. Ofitsial'nyy sayt PyTorch [Elektronnyy resurs]. Rezhim dostupa: https://pytorch.org – svobodnyy (data obrashcheniya: 17.09.2020).  [in Russian]
14. Razevig V.V. Kompleks matematicheskikh modeley i programm dlya issledovaniya i razrabotki radiogolograficheskikh sistem maloy dal'nosti: dis. … kand. tekhn. nauk. Moskva. 2016. 127 s. [in Russian]
15. Ofitsial'nyy sayt Altair FEKO [Elektronnyy resurs]. Rezhim dostupa: https://altairhyperworks.com/product/FEKO – svobodnyy (data obrashcheniya: 18.09.2020). [in Russian]
16. Bottou L., Bousquet O. The Tradeoffs of Large Scale Learning. In Sra S., Nowozin S., Wright S.J. (eds.). Optimization for Machine Learning. Cambridge: MIT Press. 2012. P. 351–368.
17. Kingma D.P., Ba J.L. Adam: A Method for Stochastic Optimization. arXiv:1412.6980. 2014.