Журнал

About Editorial Board Information for authors and reviewers Publication ethics Contacts User Account: send / review a manuscript

A harmonization technique for radar images obtained from various synthetic aperture radars to form a composite dataset for training a neural network

B.S. Savchenko^{1, 2}

¹ Moscow Institute of Physics and Technology, Dolgoprudny, Moscow Region, Russia
² JSC Racurs, Moscow, Russia

Accepted: 07.02.2024
DOI: 10.21046/2070-7401-2024-21-1-135-145

Research in the field of detection and classification of artificial objects in radar images obtained by synthetic aperture radars is one of the most promising due to the objective properties of radar data. The use of well-known correlation algorithms for object detection and classification shows limited effectiveness when working on a wide range of classes. The use of neural network algorithms, which have proven themselves well in solving a number of similar optical imaging problems, is hampered by the lack of a high-quality training sample and the specific features of radar images. The available sets of training radar data for a number of classes of interest can be obtained in a different frequency range relative to the data that are planned to be processed. The aim of the work is to confirm the possibility of joint training of a neural network on mixed radar data obtained in several frequency ranges. The paper proposes a method for forming a single training dataset consisting of real data obtained in different sensing ranges and in various combinations of polarizations. The proposed technique consists of successive stages of geometric and radiometric correction of radar images, as well as their normalization, which makes it possible to ensure the consistency of radar images of the training dataset. To confirm the effectiveness of the technique an experiment on training a neural network on raw combined data and on data that had been preprocessed in accordance with the proposed methodology was performed. The experiment showed significant improvements in the performance of the neural network while training on data that was prepared according to the proposed methodology.

Keywords: synthetic aperture radar, composite dataset, neural network, object recognition, detection and classification

Full text

References:

Antonov I. K., Detkov A. N., Nitsak D. A., Tonkikh A. N., Tsvetkov O. E., Vozdushnaya razvedka. Avtomatizirovannoe deshifrirovanie radiolokatsionnykh izobrazhenii (Aerial reconnaissance. Automated decoding of radar images), Moscow: Radiotechnics, 2021, 296 p. (in Russian).
Verba V. S., Neronsky L. B., Osipov I. G., Turuk V. E., Radiolokatsionnye sistemy zemleobzora kosmicheskogo bazirovaniya (Space-borne Earth Surveillance Radar Systems), Moscow: Radiotechnics, 2010, 680 p. (in Russian).
Vyukov N. A., Elizavetin I. V., Savchenko B. S., Sobolev A. V. (2020a), Development of methods for detecting objects in images obtained by space radars with synthesized antenna aperture, 19 th Intern. Conf. “Aviation and Cosmonautics”, Moscow: Pen, 2020, pp. 488–489 (in Russian).
Vyukov N. A., Elizavetin I. V., Savchenko B. S., Sobolev A. V. (2020b), Methods for detecting artificial objects in images obtained by space-based radars, Proc. 63 rd All-Russian Scientific Conf. MIPT, Aerospace technologies, Moscow: MIPT, 2020, pp. 18–19 (in Russian).
Turuk V. E., Verba V. S., Golovanova M. V. et al., Strizh SAR for small Condor-E satellites, Sovremennye problemy distantsionnogo zondirovaniya Zemli iz kosmosa, 2017, Vol. 14, No. 5, pp. 69–83 (in Russian), DOI: 10.21046/2070-7401-2017-14-5-69-83.
Camus B., Barbu C. L., Monteux E., Robust SAR ATR on MSTAR with Deep Learning Models trained on Full Synthetic MOCEM data, arXiv preprint, arXiv:2206.07352, 2022, DOI: 10.48550/arXiv.2206.07352.
Hui J. MAP (Mean Average Precision) for Object Detection, 2018, https://jonathan-hui.medium.com/map-mean-average-precision-for-object-detection-45c121a31173 (accessed 2.03.2023).
Inkawhich N., Inkawhich M. J., Davis E. K. et al., Bridging a gap in SAR-ATR: Training on fully synthetic and testing on measured data, IEEE J. Selected Topics in Applied Earth Observations and Remote Sensing, 2021, Vol. 14, pp. 2942–2955, DOI: 10.1109/JSTARS.2021.3059991.
Lin T. Y., Dollár P., Girshick R. et al., Feature pyramid networks for object detection Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2017, pp. 2117–2125, DOI: 10.48550/arXiv.1612.03144.
Liu Z., Luo S., Wang Y., Mix MSTAR: A Synthetic Benchmark Dataset for Multi-Class Rotation Vehicle Detection in Large-Scale SAR Images, Remote Sensing, 2023, Vol. 15, No. 18, Article 4558, DOI: 10.3390/rs15184558.
Wang H., Chen S., Xu F., Jin Y. Q., Application of deep-learning algorithms to MSTAR data, 2015 IEEE Intern. Geoscience and Remote Sensing Symp. (IGARSS), 2015, pp. 3743–3745, DOI: 10.1109/IGARSS.2015.7326637.
Wu Y., Kirillov A., Massa F., Lo W. Y., Girshick R., Detectron2, 2019, https://github.com/facebookresearch/detectron2 (accessed 20.10.2023).

создание сайта - TXL