Comparative analysis of the efficiency of OBB and HBB detectors with an additional classification stage based on neural network architectures YOLO
Year: 2026
Type: article
Abstract: This paper considers the problem of detecting and classifying ground objects in radar images using deep learning methods. The study compares two architectures: a single-stage architecture based on the YOLO detector and a two-stage architecture combining YOLO with an additional classifier. The experiments evaluate YOLOv5 with horizontal bounding boxes (HBB) and YOLOv11 with oriented bounding boxes (OBB) as object detectors. In the two-stage system, the study uses an additionally trained ConvNeXt_tiny classifier to solve the binary “object/background” classification task. The experiments use the developed binMSTAR dataset, which is based on the mixMSTAR dataset. The results demonstrate that the two-stage system does not improve the final detection quality, but only redistributes errors between false positives and missed detections. The single-stage models provide balanced Precision and Recall values: 0.967 and 0.965 for YOLOv5, and 0.970 and 0.963 for YOLOv11, respectively. The study concludes that the single-stage YOLO detectors are suitable for real-time binary detection of ground objects in radar images, where all detected targets belong to the single “object” class.
Source: Ural Radio Engineering Journal
Authors: D. D. Ivanov, V. I. Weber, I. M. Madaliev
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7234762860
