Homepage Program Systems: Theory and Applications Русская версия
ISSN 2079-3316 Bilingual online scientific Online scientific journal of the Ailamazyan Program System Institute of the Ailamazyan PSI of PSI of Russian Academy of Science of RAS 12+ 
Volume 16 (2025) . Issue 4 (67) . Paper No. 7 (454)

Artificial intelligence and machine learning

Research Article
DOIDigital Object Identifier 10.25209/2079-3316-2025-16-4-173-216

Comparative Analysis of Backbone Architectures for Instance Segmentation of Objects in Aerial Imagery Using Mask R-CNN

Igor Victorovich Vinokurov1Correspondent author, Daria Aleksandrovna Frolova2, Andrey Ivanovich Ilyin3, Ivan Romanovich Kuznetsov4

1-4Financial University under the Government of the Russian Federation, Moscow, Russia
1 Igor Victorovich Vinokurov — Correspondent author igvvinokurov@fa.ru

Abstract. This paper compares Mask R-CNN models with various pretrained backbone architectures for implementing instance segmentation of real estate objects in aerial images. The models were fine-tuned on a specialized dataset provided by the PLC « Roskadastr».

Analysis of the accuracy of detecting bounding boxes and object segmentation masks revealed the preferred architectures: Swin transformers (Swin-S and Swin-T) and the ConvNeXt-T convolutional network. The high accuracy of these models is explained by their ability to account for global contextual dependencies of the image.

The results of the study allow us to formulate the following recommendations for choosing a backbone architecture: for real-time monitoring systems where performance is critical, lightweight models (EfficientNet-B3, ConvNeXt-T, Swin-T) are advisable; for offline tasks requiring maximum accuracy (such as real estate mapping), the large-scale Swin-S model is recommended. (Linked article texts in English and in Russian).

Keywords: instance segmentation, backbone, Mask R-CNN, ResNet, DenseNet, EfficientNet, ConvNeXt, Swin

MSC-20202020 Mathematics Subject Classification 68T20; 68T07, 68T45MSC-2020 68-XX: Computer science MSC-2020 68Txx: Artificial intelligence MSC-2020 68T20: Problem solving in the context of artificial intelligence (heuristics, search strategies, etc.) MSC-2020 68T07: Artificial neural networks and deep learning MSC-2020 68T45: Machine vision and scene understanding

For citation: Igor V. Vinokurov, Daria A. Frolova, Andrey I. Ilyin, Ivan R. Kuznetsov. Comparative Analysis of Backbone Architectures for Instance Segmentation of Objects in Aerial Imagery Using Mask R-CNN. Program Systems: Theory and Applications, 2025, 16:4, pp. 173–216. (in Engl. In Russ.). https://psta.psiras.ru/2025/4_173-216.

Full text of bilingual article (PDF): https://psta.psiras.ru/read/psta2025_4_173-216.pdf (Clicking on the flag in the header switches the page language).

The article was submitted 22.09.2025; approved after reviewing 27.09.2025; accepted for publication 12.10.2025; published online 19.10.2025.

© Vinokurov I. V., Frolova D. A., Ilyin A. I., Kuznetsov I. R.
2025
Editorial address: Ailamazyan Program Systems Institute of the Russian Academy of Sciences, Peter the First Street 4«a», Veskovo village, Pereslavl area, Yaroslavl region, 152021 Russia;   Website:  http://psta.psiras.ru Phone: +7(4852) 695-228;   E-mail: ;   License: CC-BY-4.0License text on the Creative Commons site
© Ailamazyan Program System Institute of Russian Academy of Science (site design) 2010–2025