B.S. Goryachkin1, N.S. Podoprigorova2, S.S. Podoprigorova3
1–3 Bauman Moscow State Technical University (Moscow, Russia)

1 bsgor@mail.ru, 2 n.podoprigorova@icloud.com, 3 s.podoprigorova@icloud.com

Finding an effective neural network architecture takes a lot of time and requires significant computational resources. The goal of the article is to establish a relationship between the architecture features and its output characteristics so that, even before training the model, we can separate high-performance architectures from suboptimal ones.

The article analyzes how the neural network architecture affects the model characteristics: training time, processing speed, and model size. The experiments have been conducted on a model based on the U-Net architecture. The main model characteristics have been changed: depth, width (average filter size), size of convolution kernels, and the presence or absence of skip connections. The problem of semantic segmentation of forest cover violations in a data set with satellite images has been solved.

The established dependencies should facilitate the process of finding an effective neural network architecture for the semantic segmentation task.

Goryachkin B.S., Podoprigorova N.S., Podoprigorova S.S. Ergonomic analysis of forest cover disturbance mapping methods using neural networks. Neurocomputers. 2025. V. 27. № 2. P. 32–42. DOI: https://doi.org/10.18127/j19998554-202502-04 (in Russian)

1. Zoph B., Le Q.V. Neural architecture search with reinforcement learning. Proc. of the 5th International Conference on Learning Representations. 2017.
2. Lopes V. et al. EPE-NAS: Efficient performance estimation without training for neural architecture search. International Conference on Artificial Neural Networks. 2021. P. 552–563.
3. Shen X. et al. DeepMAD: Mathematical architecture design for deep convolutional neural network. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver, BC, Canada. 2023. P. 6163–6173. DOI: 10.1109/CVPR52729.2023.00597.
4. Forrest I. Exploring the design space of deep convolutional neural networks at large scale. Dissertation. University of California, Berkeley. 2016. URL: https://doi.org/10.48550/arXiv.1612.06519.
5. Zebin T., Scully P.J., Peek N., Casson A.J., Ozanyan K.B. Design and implementation of a convolutional neural network on an edge computing smartphone for human activity recognition. IEEE Access. 2019. V. 7. P. 133509–133520. DOI: 10.1109/ACCESS.2019.2941836.
6. Shikhov A.N., Chernokul'skij A.V., Azhigov I.O. Prostranstvenno-vremennoe raspredelenie vetrovalov v lesnoj zone Zapadnoj Sibiri v 2001–2020 gg. Sovremennye problemy distantsionnogo zondirovaniya Zemli iz kosmosa. 2022. T. 19. № 3. S. 186–202. DOI: 10.21046/ 2070-7401-2022-19-3-186-202. (in Russian)
7. Tarasov A.V., Shikhov A.N., Shabalina T.V. Raspoznavanie narushenij lesnogo pokrova po sputnikovym snimkam Sentinel-2 s pomoshch'yu svertochnykh nejronnykh setej. Sovremennye problemy distantsionnogo zondirovaniya Zemli iz kosmosa. 2021. T. 18. № 3. S. 51–64. DOI: 10.21046/2070-7401-2021-18-3-51-64.
8. Patel C., Bhatt D., Sharma U. et al. DBGC: Dimension-based generic convolution block for object recognition. Sensors. 2022. V. 22. № 1780. DOI: https://doi.org/10.3390/s22051780.
9. Sankararaman K.A., De S., Xu Zh. et al. The impact of neural network overparameterization on gradient confusion and stochastic gradient descent. Proceedings of the 37th International Conference on Machine Learning. Vienna, Austria. 2020.