Small object detection in radar images

Weber V.I., Ivanod D.D., Kuprits V.Y., Svetlichny Y.A.
 pdf (4495K)

This study examines how object annotation format and input image resolution jointly affect the quality of binary detection of small objects in radar images. The study aims to compare YOLO family models, including YOLOv5, YOLOv8, YOLOv11, and YOLOv26, with specialized oriented object detection architectures such as Rotated RetinaNet, R3Det, ROI Transformer, Gliding Vertex, and related models in the task of binary detection of small ground objects in radar imagery. The study also aims to identify data preparation parameters that improve object localization accuracy.

For the experimental analysis, we formed a new binMSTAR dataset from the open mixMSTAR database. The dataset contains annotations in horizontal bounding box (HBB) and oriented bounding box (OBB) formats for nine input frame resolutions ranging from 640 × 640 to 1408 × 1408 pixels. We assessed the statistical significance of the observed differences using the nonparametric Mann – Whitney U test. The results show that OBB annotation provides a statistically significant advantage in terms of mAP50–95, with an average improvement of 6.5%, because it reduces the influence of redundant background noise. The experiments also reveal an optimal input frame size range of 896–1024 pixels. Further resolution growth decreases localization accuracy, since excessive background content weakens the object-related feature representation. YOLO models demonstrate the best overall balance of detection quality, with YOLOv11(m) reaching a peak mAP50–95 of 0.87 at a resolution of 960 pixels. At the same time, specialized OBB detectors at a resolution of 1024 pixels achieve comparable recall values, reaching up to 0.997. These results support the use of oriented bounding boxes and input resolutions within the 896–1024 pixel range for binary detection tasks in radar imagery.

Keywords: radar image, synthetic aperture radar, object detection, YOLO models, small object, binMSTAR, mixMSTAR, horizontal bounding boxes, oriented bounding boxes, input resolution, deep learning
Citation in English: Weber V.I., Ivanod D.D., Kuprits V.Y., Svetlichny Y.A. Small object detection in radar images // Computer Research and Modeling, 2026, vol. 18, no. 5, pp. 1249-1268
Citation in English: Weber V.I., Ivanod D.D., Kuprits V.Y., Svetlichny Y.A. Small object detection in radar images // Computer Research and Modeling, 2026, vol. 18, no. 5, pp. 1249-1268
DOI: 10.20537/2076-7633-2026-18-5-1249-1268
URL: http://crm-en.ics.org.ru/journal/article/3779/
Creative Commons License This work is licensed under a Creative Commons Attribution-NoDerivs 3.0 Unported License.

Copyright © 2026 Weber V.I., Ivanod D.D., Kuprits V.Y., Svetlichny Y.A.

Indexed in Scopus

Full-text version of the journal is also available on the web site of the scientific electronic library eLIBRARY.RU

The journal is included in the Russian Science Citation Index

The journal is included in the RSCI

International Interdisciplinary Conference "Mathematics. Computing. Education"

Website Copyright © 2009–2026 Institute of Computer Science