Method for Recognizing Objects in Thermal Images
Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì · 22.06.2026 · 19 · 213-229
DOI: 10.23939/sisn2026.19.213 ↗Kurz gefasst
Fachartikel in der Zeitschrift Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì, online erschienen am 17.05.2026. Verfasst von Ivan Peleshchak und 3 weiteren Autor:innen. Die Arbeit nennt 52 Literaturangaben.
Abstract (Originalsprache)
The paper addresses the problem of improving the accuracy of automated object detection in thermal images under conditions of low contrast, sensor noise, and structural uncertainty of the terrain. The relevance of this research is driven by the increasing use of unmanned aerial vehicles for monitoring and reconnaissance, where thermal imaging systems serve as a key source of information in low-visibility environments. Standard convolutional neural network architectures, particularly detectors of the YOLO family, demonstrate high processing speed; however, their effectiveness decreases when applied to infrared imagery due to the specific characteristics of thermal signatures and the presence of impulse noise. This work proposes a method of spectral feature aggregation based on Singular Value Decomposition (SVD), integrated into the pooling mechanism of the YOLOv8 neural network. Unlike traditional Max Pooling and Average Pooling operations, the developed SVD pooling performs local spectral filtering of feature maps by extracting the largest singular component corresponding to the structurally significant part of the signal. This approach enables effective suppression of high-frequency thermal noise while preserving the spatial patterns of objects. A modified YOLOv8-EPB-Lite-SVD architecture based on EfficientNet-B0 and BiFPN with integrated spectral filtering modules has been developed. Experimental verification of the model was conducted on a dataset of thermal images acquired from unmanned platforms. The results demonstrate that the integration of SVD pooling increases the Precision metric to 0.881 and achieves an mAP@50 value of 0.852 while maintaining the number of model parameters at 4.56 million. The obtained results are statistically equivalent to those of “heavy” architectures with four times more parameters. The influence of numerical precision on algorithm stability was also investigated. It was shown that the use of full precision (FP32) is critically necessary to prevent the loss of significance of small singular values under mixed-precision (FP16) computation.
Zitieren
APA
Peleshchak, I., Puchkov, M., Peleshchak, R., & Synko, A. (2026). Method for Recognizing Objects in Thermal Images. Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì, 19, 213-229. https://doi.org/10.23939/sisn2026.19.213
Vancouver
Peleshchak I, Puchkov M, Peleshchak R, Synko A. Method for Recognizing Objects in Thermal Images. Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì. 2026;19:213-229. doi:10.23939/sisn2026.19.213
BibTeX
@article{Peleshchak2026,
author = {Peleshchak, Ivan and Puchkov, Maksym and Peleshchak, Roman and Synko, Anna},
title = {Method for Recognizing Objects in Thermal Images},
journal = {Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì},
year = {2026},
volume = {19},
pages = {213--229},
doi = {10.23939/sisn2026.19.213}
}
EUVIMED