D.E. Boriskin1, N.V. Gorbachev2, I.D. Isaev3, A.N. Savelyev4

1-4Bauman Moscow State Technical University (national research university) (Moscow, Russia)

1bde19k023@student.bmstu.ru, 2gorbachevnv@student.bmstu.ru, 3isaevid@bmstu.ru, 4savelyev.an@bmstu.ru

Software superresolution methods for their effective application require the presence of a priori information about the signal-noise formation in an analyzed digital two-coordinate radar image frame, which differs significantly in its various fragments. The classic way to overcome the absence or lack of a priori information used in well-known basic methods is to implement iterative algorithms which step-by-step approximate the solution to requirements set by the objective function. At the same time, it’s advisable to adapt regularization parameters of algorithms used to the observation conditions in various fragments of digital radar image frame.

The problem of parameters adapting of image reconstruction algorithms to radar surveillance conditions can be solved by training of algorithms. Currently, along with the basic ones, convolutional neural network (CNN) algorithms are widely used, which used, among other things, to solve super-resolution problems. CNN performance in accordance with specified criterion is determined by its structure, training and signal-to-noise ratio assessment methods.

Simulation and comparative analysis of basic superresolution methods (inverse filtering, Wiener filtration, Tikhonov method, Lucy-Richardson method, Split Bregman Algorithm) and modifications of convolutional neural network of the U-Net architecture are performed. For numerical experiments, a generator forming digital radar image frames with random parameters as well as noise power estimation module have been developed. The stability of convolutional neural network algorithm at low signal-to-noise ratios has been established in comparison with classical algorithms.

Preferred areas of SNR for basic and convolutional neural network algorithms for object superresolution on synthesized digital radar image frames with random parameters have been identified: among traditional methods, the Wiener filter showed the best results.

The negative phenomenon of neural network "hallucinations" for SNR less than 0 dB with a limitation of its use in these conditions has been revealed.

For an objective efficiency assessment, it is proposed to further evaluate resource costs required to compare algorithms, such as time, computational and economic costs.

1. Bogomolov Yu.V., Aristov V.V., Vinogradov A.V. i dr. Obzor metodov obratnoi svertki. Uspekhi fizicheskikh nauk. 2023. T. 193. № 6. S. 669−685. DOI: https://doi.org/10.3367/UFNr.2022.05.039189. (in Russian)
2. Zhang Q., Wu R., Zhang Y. i dr. A Bayesian Super-Resolution Method for Forward-Looking Scanning Radar Imaging Based on Split Bregman. Proceedings of IGARSS 2018. 2018 IEEE International Geoscience and Remote Sensing Symposium. 22−27 July 2018. Valencia, Spain. 2018. P. 5135−5138. DOI: 10.1109/IGARSS.2018.8518359.
3. Zhang Y. et al. Fast Split Bregman Based Deconvolution Algorithm for Airborne Radar Imaging. Remote Sensing. MDPI AG. 2020. V. 12. № 11. P. 1747.
4. Martinello M., Favaro P. Single Image Blind Deconvolution with HigherOrder Texture Statistics. Video Processing and Computational Video: International Seminar. Dagstuhl Castle, Germany. October 10−15 2010. Revised Papers. D. Cremers et all (editors). Berlin. Heidelberg: Springer Berlin Heidelberg. 2011. S. 124−151. ISBN 978-3-642-24870-2. DOI:10.1007/978-3-642-24870-2_6. https://doi.org/10.1007/978-3-642-24870-2_6.
5. Carasso A., Bright D., Vladar A. The APEX Method and Real-Time Blind Deconvolution of Scanning Electron Microscope Imagery. 2001-11-01-00:11:00/2001. DOI: https://doi.org/10.6028/NIST.IR.6835. https://tsapps.nist.gov/publication/get pdf.cfm?pub id=50987.
6. Yi C., Shimamura T. An Improved Maximum-Likelihood Estimation Algorithm for Blind Image Deconvolution Based on Noise Variance Estimation. Journal of Signal Processing. 2012. V. 16. № 6. P. 629−635. DOI: 10.2299/jsp.16.629.
7. Maik V., Aishwarya R., Paik J. Blind deconvolution using maximum aposteriori (MAP) estimation with directional edge based priori. Optik. 2018. V. 157. P. 1129−1142. DOI: 10.2299/jsp.16.629.
8. Justen L., Ramlau R. A non-iterative regularization approach to blind deconvolution. Inverse Problems. April 2006. V. 22. № 3. P. 771−800. DOI: 10.1088/0266-5611/22/3/003. https://doi.org/10.1088/0266-5611/22/3/003.
9. Justen L., Ramlau R. Extensions of the Justen-Ramlau blind deconvolution method. Advances in Computational Mathematics. 2013. V. 39. P. 465−491. DOI: 10.1007/s10444-012-9290-z. https://doi.org/10.1007/s10444-012-9290-z.
10. Wiener N. Extrapolation, interpolation and smoothing of stationary time series, with engineering applications. Technology Press of the Massachusetts Institute of Technology. 1957. 163 p.
11. Dong J., Roth S., Schiele B. Deep wiener deconvolution: Wiener meets deep learning for image deblurring. Advances in Neural Information Processing Systems. 2020. V. 33. P. 1048−1059.
12. Barakat V. i dr. Model-Based Tikhonov-Miller Image Restoration. Proceedings of the 1997 International Conference on Image Processing (ICIP 97). 3 Volume Set. V. 1. USA: IEEE Computer Society. 1997. S. 310. (ICIP ’97). ISBN 0818681837.
13. Liu H. et al. Spectral deconvolution and feature extraction with robust adaptive Tikhonov regularization. IEEE Transactions on Instrumentation and Measurement. 2012. V. 62. № 2. P. 315−327.
14. Arsenin V.Ya., Tikhonov A.N. Tikhonovskaya regulyarizatsiya. Bolshaya rossiiskaya entsiklopediya: nauchno-obrazovatelnyi portal. https://bigenc.ru/c/tikhonovskaia-reguliarizatsiia-cbcb56/?v=3786693 (08.06.2022). (in Russian)
15. Fish D.A. i dr. Blind deconvolution by means of the Richardson–Lucy algorithm. J. Opt. Soc. Am. A. January 1995. V. 12. № 1. P. 58−65. DOI: 10.1364/JOSAA.12.000058. http://opg.optica.org/josaa/abstract.cfm?URI=josaa-12-1-58.
16. Francisco J. Ávila i dr. Iterative-Trained Semi-Blind Deconvolution Algorithm to Compensate Straylight in Retinal Images. Journal of Imaging. 2021. V. 7. № 4. DOI: 10.3390/jimaging7040073. https://www.mdpi.com/2313-433X/7/4/73.
17. Lucy L.B. An iterative technique for the rectification of observed distributions. Astronomical Journal. 1974. V. 79. № 6. P. 745−754.
18. Panfilova K., Umnyashkin S. Linear blur compensation in digital images using Lucy-Richardson method. IEEE EIConRusNW. 2015. P. 163−167.
19. Goldstein T., Osher S. The Split Bregman Method for L1-Regularized Problems. SIAM Journal on Imaging Sciences. 2009. V. 2. № 2. P. 323−343.
20. Tatuzov A.L. Neironnye seti v zadachakh radiolokatsii. Kn. 28. M.: Radiotekhnika. 2009. 432 s. (Nauchnaya seriya «Neirokompyutery i ikh primenenie»). (in Russian)
21. Galushkin A.I. Neironnye seti. Bolshaya rossiiskaya entsiklopediya: nauchno-obrazovatelnyi portal. https://bigenc.ru/c/neironnye-seti-e734b3/?v=9558530 (26.05.2025). Data publikatsii: 16.11.2022. Data obnovleniya: 25.01.2024. (in Russian)
22. Nair V., Hinton G. Rectified Linear Units Improve Restricted Boltzmann Machines. Proceedings of the 27th International Conference on Machine Learning (ICML-10). Haifa, Israel. 21−24 June 2010. V. 27. P. 807−814.
23. Ronneberger O., Fischer P., Brox T.O. U-Net: Convolutional Networks for Biomedical Image Segmentation. 2015. https://arxiv.org/abs/1505.04597.
24. Kingma D.P., Ba J. Adam: A Method for Stochastic Optimization. 2017. https://arxiv.org/abs/1412.6980 (26.05.2025).
25. Giperparametr. Bolshaya rossiiskaya entsiklopediya: nauchno-obrazovatelnyi portal.
    https://bigenc.ru/c/giperparametr-b6e604/?v=7929116 (26.07.2023). (in Russian)
26. Goldstein T., Osher S. The Split Bregman Method for L1-Regularized Problems. SIAM Journal on Imaging Sciences. 2009. V. 2. № 2. P. 323−343. DOI: 10.1137/080725891.