Abstract. A system is being developed to counter UAVs and drones by monitoring the airspace and identifying aerial objects and their flight paths. The main component of the proposed system is a UAV identification unit that classifies photographic images using an embedded computer vision neural network. The problem of designing this neural network for ground-based observation of aerial objects in the optical band is considered in order to determine the main geometric parameters of these objects for their subsequent classification and the formation of corresponding information signals and features at the network output. These features are expected to be used as input data for analyzing behavioral algorithms and generating the trajectory of the observed objects. A neural network architecture optimized for the problem at hand and consisting of convolutional and fully connected layers has been designed. To obtain a more flexible logical inference, the program implements elements of fuzzy logic. Various source image augmentation methods based on Gaussian blur filters, as well as rotation and reflection algorithms, are implemented using a stochastic algorithm. Special attention is paid to neural network training algorithms, solving optimization issues, and adapting the training process to the problem being solved. A method based on the Adam optimizer is proposed for adaptively adjusting the gradient descent step size at each calculation stage within the considered training batch, taking into account the change history of the activation function and cost function gradients. A computer program for neural network deployment and training has been developed. Multi-stage training and testing of the program on images of aerial objects from seven different classes were successfully conducted. Special attention was paid to the reproducibility of the training results. Along with the final classification accuracy, the stability of the probability distribution across classes was monitored, as well as the repeatability of the results during repeated training runs on identical datasets.

Keywords: computer vision, convolutional neural network, parallel computing, cross-entropy loss, regularization, Adam optimizer, activation function, fuzzy logic

For citation: Molchanov D. A., Tretyakov A. V. Aerial object identification system using computer vision. Aerospace Engineering and Technology. 2026. Vol. 4, no. 1, pp. 124–143. (In Russ.) https://doi.org/10.52467/2949-401X-2026-4-1-124-143; https://elibrary.ru/QCYIHU

References
1. Gusejnov A.B, Lyapunov V.V., Trusov V.N. Proektirovanie sistem upravleniya krylatyh raket [Design of cruise missile control systems]. Moscow.: MAI, 2020, 192 p. (In Russ.).
2. Tret’yakov A.V., Piskunova O.I., Tarasenko A.V. Proektirovanie sistem avtomaticheskogo upravleniya dlya aviacionnyh ustrojstv [Design of automatic control systems for aviation devices]. Dubna: Dubna University, 2025. 221 p. (In Russ.)
3. Liu L., Jiang H., He P., Chen W., Liu X., Gao J., Han J. On the Variance of the Adaptive Learning Rate and Beyond. Proceedings of the Eighth International Conference on Learning Representations (ICLR 2020). URL: https://arxiv.org/abs/1908.03265 (accessed: March 18, 2026).
4. G. De Luca, Aibin M. Neural Network Architecture: Criteria for Choosing the Number and Size of Hidden Layers. Baeldung on Computer Science. 2025. URL: https://translated.turbopages.org/proxy_u/en-ru.ru.3fc11e68-69c3d5ae-aebd9d0e-74722d776562/https/www.baeldung.com/cs/neural-networks-hidden-layers-criteria (accessed: March 18, 2026).
5. Terehov V.I., Chernenkiy I.M., Minakova S.V. Research and development of methods for adjustment the number of neurons in the hidden layer of fully connected neural network. Nejrokomp’yutery: razrabotka, primenenie [Journal Neurocomputers]. 2017. No. 3, pp. 52–62. (In Russ.)
6. Frolov V.V., Zholtkevych G.N., Prihodko O.Yu., Skuryatin Yu.V. Method of preliminary selection of multilayer neural network architecture for polyharmonic signal approximation // Vestnik of RSREU. 2021. No 78, pp. 102–111. (In Russ.). https://doi.org/10.21667/1995-4565-2021-78-102-111
7. Natalenko D. N. Re-Identification of Wild Animals in Images Using Deep Learning Methods. Vestnik komp’iuternykh i informatsionnykh tekhnologii [Herald of computer and information technologies]. 2026. No. 2. pp. 13-22. (In Russ.) https://doi.org/10.14489/vkit.2026.02.pp.013-022
8. Kaliyev D.I., Shvets O.Ya. Convolutional neural networks for solving fire detection problems based on aerial photography. Program sistem: theory and applications. 2022. Vol. 13, no. 1(52). pp. 195-213. (In Russ.). https://doi.org/10.25209/2079-3316-2022-13-1-195-213
9. Bentz J. CUDA 13.2 Introduces Enhanced CUDA Tile Support and New Python Features. NVIDIA Technical Blog. 9.03.2026. URL: https://developer.nvidia.com/blog/cuda-13-2-introduces-enhanced-cuda-tile-support-and-new-python-features/ (accessed: March 22, 2026).
10. Vakulenko S.A., Zhihareva A.A. Prakticheskij kurs po nejronnym setyam [A practical course on neural networks]. Saint Petersburg: IТМО University, 2018. 71 p. (In Russ.).
11. Vaskan V.D. Overview of convolutional neural network architectures for the image classification problem. IT-Standard. 2021. No. 3. URL: https://itstd-journal.ru/?page_id=1080&article=234 (accessed: December 03, 2025). (In Russ.).
12. Kingma D. P., Ba J. L. Adam: A Method for stochastic Optimization // Proceedings of the Conference Track 3rd International Conference on Learning Representations, (ICLR), San Diego, CA, USA, May 7-9, 2015. URL: https://doi.org/10.48550/arXiv.1412.6980 (accessed: March 23, 2026).
13. Liu H., Jia X., Su C., Yang H., Li C. Tire appearance defect detection method via combining HOG and LBP features. Frontiers in Physics. 2023. Vol. 10. 12 p. https://doi.org/10.3389/fphy.2022.1099261
14. Dalal N., Triggs B. Histograms of Oriented Gradients for Human Detection // Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), San Diego, CA, USA, 20-25 June 2005. IEEE, 2025. Vol. 1. pp. 886-893. https://doi.org/10.1109/CVPR.2005.177
15. Mumuni A., Mumuni F., Gerrar N.K. A survey of synthetic data augmentation methods in computer vision. Machine Intelligence Research. 2024. Vol. 21. Pp. 831–869. https://doi.org/10.1007/s11633-022-1411-7
16. Shorten C., Khoshgoftaar T.M. A survey on Image Data Augmentation for Deep Learning // Journal of Big Data. 2019. Vol. 6, no.1. Article 60. https://doi.org/10.1186/s40537-019-0197-0
17. Johnson J., Alahi A., Fei-Fei L. Perceptual Losses for Real-Time Style Transfer and Super-Resolution // Computer Vision – ECCV 2016. ECCV 2016. Lecture Notes in Computer Science, vol 9906. Springer, Cham., 2016. https://doi.org/10.1007/978-3-319-46475-6_43
18. Borovik I.G., Nguуen K.T. A comparison of softmax and sigmoid activation functions in the context of multiclass object recognition classification. [International Scientific Research Journal] 2024. No. 6 (144). 6 p. (In Russ.) https://doi.org/10.60797/IRJ.2024.144.174