E.A. Samoylin1, S.V. Tatarintsev2, D.Yu. Dronov3, K.E. Skugorov4

1 Military Training and Research Center of the Air Force 

«Air Force Academy named after Professor N.E. Zhukovsky and Y.A. Gagarin»  of the Russian Federation Ministry of Defence (Voronezh, Russia)

At present, when receiving and processing digital images, they often encounter cases of various noise and distortions appearing on the raster. In practice, we most often have to deal with applicative (impulsive) and additive Gaussian noise. At the same time, noise with a brightness distribution close to a truncated Gaussian, the mode of which falls on the upper or lower limits of the image brightness quantization, can be attributed to a separate class of applicative distortions, which are most often encountered in practice.  At present, almost all known algorithms for correcting applicative distortions in images are spatially selective, when at the first stage of processing, the detection of distorted image elements is performed, and at the second stage, the restoration of lost (due to distortions) image elements by some method. To date, a sufficient number of algorithms for detecting (as a rule, statistical) such applicative distortions have been proposed; however, a priori unknown of the laws and parameters of the distribution of random signals and noise, as well as a priori probabilities of the presence and absence of distortions in the image, does not allow achieving their potential quality. One of the ways out of this situation can be the use of a neural network approach and its combination with statistically optimal algorithms. In the problem of detecting distortions, obtaining the necessary information about such parameters of distortions as the prior probability of their presence in the image and the standard deviation is possible through the use of artificial neural networks with high approximation properties acquired by them in the course of training on a training set of examples. This information can then be used by a statistically optimal algorithm. 

In accordance with this, the aim of the work is to develop a detector of applicative distortions in digital images based on a combination of a neural and statistically optimal algorithm. 

The essence of the proposed algorithm is to use at the first stage of processing a trained neural network capable of functioning under conditions of complete a priori uncertainty and allowing to obtain approximate estimates of a priori information about distortions in the image, and at the second stage of processing – rules based on the criterion of the minimum average risk (Bayesian criterion) using neural network estimates.

As shown by the results of numerical studies, the proposed approach combines the advantage of neural networks, which consists in their high approximating properties, acquired in the course of training on precedents, as well as the optimality of a statistical detector. The presented results of numerical studies of the efficiency of the proposed detector indicate its advantage over the known detector in almost the entire possible range of applicative interference intensity.

Samoylin E.A., Tatarintsev S.V., Dronov D.Yu., Skugorov K.E. Detection of applicative distortions in digital images based on a combination of a neural network and a statistically optimal algorithm. Neurocomputers. 2021. V. 23. № 5. Р. 21−32. DOI: https://doi.org/10.18127/j19998554-202105-02 (In Russian).

1. Gonsales R., Vuds R. Tsifrovaya obrabotka izobrazheniy. M.: Tekhnosfera. 2012. 1104 s. (in Russian).
2. Khuang T.S., Eklund Dzh.-O., Nussbaumer G.Dzh., Zokhar Sh., Yustusson B.I., Tyan Sh.-G. Bystryye algoritmy v tsifrovoy obrabotke izobrazheniy. M.: Radio i svyaz. 1984. 224 s. (in Russian).
3. Gruzman I.S., Kirichuk V.S., Kosykh V.P., Peretyagin G.I., Spektor A.A. Tsifrovaya obrabotka izobrazheniy v informatsionnykh sistemakh: Uchebnoye posobiye. Novosibirsk: Izd-vo NGTU. 2000. 168 s. (in Russian).
4. Yukhno P.M. Prednamerennyye opticheskiye pomekhi vysokotochnomu oruzhiyu. M.: Radiotekhnika. 2017. 640 s. (in Russian).
5. Samoylin E.A. Prostranstvenno-izbiratelnaya filtratsiya polutonovykh izobrazheniy. Radiotekhnika. 2006. № 12. S. 42–45. (in Russian).
6. Samoylin E.A., Slavnov K.V., Tatarintsev S.V. Sintez algoritma raboty statisticheski optimalnogo obnaruzhitelya modalnykh applikativnykh pomekh s usechennym gaussovskim raspredeleniyem yarkosti na tsifrovykh opticheskikh izobrazheniyakh. Radiotekhnika. 2021. T. 85. № 1. S. 96–103. (in Russian).
7. Samoylin E.A. Sintez statisticheski optimalnykh obnaruzhiteley impulsnykh pomekh pri filtratsii izobrazheniy. Radiotekhnika. 2009. № 9. S. 25–31. (in Russian).
8. Stepanov O.A. Neyrosetevyye algoritmy v zadache nelineynogo otsenivaniya. Vzaimosvyaz s bayesovskim podkhodom. Sb. materialov KhI konf. molodykh uchenykh. S-Pb.: GNTs OAO Kontsern TsNII «Elektropribor». 2009. S. 39–65. (in Russian).
9. Kirsanov E.A., Sirota A.A. Obrabotka informatsii v prostranstvenno-raspredelennykh sistemakh radiomonitoringa: statisticheskiy i neyrosetevoy podkhody. M.: FIZMATLIT. 2012. 344 s. (in Russian).
10. Uossermen F. Neyrokompyuternaya tekhnika: teoriya i praktika. M.: Mir. 1992. 240 s. (in Russian).
11. Levin B.R. Teoreticheskiye osnovy statisticheskoy radiotekhniki. Kn. 1. M.: Sov. radio. 1969. 752 s. (in Russian).
12. Samoylin E.A. Povysheniye stepeni funktsionalnoy polnoty neyropodobnykh elementov modeli neyronnoy seti. Neyrokompyutery: razrabotka. primeneniye. 2010. № 10. S. 17–26. (in Russian).