@article{ART003397466},
author={Seonghak Kim},
title={Ontology-Guided Data Engine Framework for Defense Domain Data Construction},
journal={Journal of The Korea Society of Computer and Information},
issn={1598-849X},
year={2026},
volume={31},
number={10},
pages={269-280}
TY - JOUR
AU - Seonghak Kim
TI - Ontology-Guided Data Engine Framework for Defense Domain Data Construction
JO - Journal of The Korea Society of Computer and Information
PY - 2026
VL - 31
IS - 10
PB - The Korean Society Of Computer And Information
SP - 269
EP - 280
SN - 1598-849X
AB - The performance of defense Artificial Intelligence (AI) systems relies heavily on the acquisition of high-quality training data; however, constructing training datasets for rare military objects―such as enemy equipment and new weapon systems―poses significant challenges due to security constraints and the inherent scarcity of such data. To address this data scarcity problem, this paper proposes an Ontology-Guided Data Engine, an automated data construction framework that integrates the ontological structure of Wikidata with large vision foundation models. The proposed framework defines hierarchical concepts of military assets and utilizes visual prototypes from reference images to semantically filter valid data from the web, open datasets, and internal archives. Subsequently, it employs Grounding DINO and SAM to automatically generate precise bounding boxes and segmentation masks in the COCO format with minimal human intervention. Quantitative evaluations on 12 military classes demonstrate that the proposed engine achieves a data selection precision of 0.720 with a micro F1 of 0.729, an auto-labeling quality of 82.6% Box mAP and 83.9% Mask mAP, and the highest downstream detection and segmentation performance (54.9% Box mAP, 51.7% Mask mAP), outperforming keyword crawling, CLIP-only filtering, and SAM 3 baselines. These results confirm that the framework effectively identifies fine-grained differences in weapon systems that text-prompt-based models often fail to distinguish. Consequently, this framework is anticipated to serve as an efficient methodology for future research on Data-Centric AI within the defense domain.
KW - Defense AI;Data Scarcity;Data Mining;Ontology-Guided Data Engine;Auto-Labeling;Vision Foundation Model
DO -
UR -
ER -
Seonghak Kim. (2026). Ontology-Guided Data Engine Framework for Defense Domain Data Construction. Journal of The Korea Society of Computer and Information, 31(10), 269-280.
Seonghak Kim. 2026, "Ontology-Guided Data Engine Framework for Defense Domain Data Construction", Journal of The Korea Society of Computer and Information, vol.31, no.10 pp.269-280.
Seonghak Kim "Ontology-Guided Data Engine Framework for Defense Domain Data Construction" Journal of The Korea Society of Computer and Information 31.10 pp.269-280 (2026) : 269.
Seonghak Kim. Ontology-Guided Data Engine Framework for Defense Domain Data Construction. 2026; 31(10), 269-280.
Seonghak Kim. "Ontology-Guided Data Engine Framework for Defense Domain Data Construction" Journal of The Korea Society of Computer and Information 31, no.10 (2026) : 269-280.
Seonghak Kim. Ontology-Guided Data Engine Framework for Defense Domain Data Construction. Journal of The Korea Society of Computer and Information, 31(10), 269-280.
Seonghak Kim. Ontology-Guided Data Engine Framework for Defense Domain Data Construction. Journal of The Korea Society of Computer and Information. 2026; 31(10) 269-280.
Seonghak Kim. Ontology-Guided Data Engine Framework for Defense Domain Data Construction. 2026; 31(10), 269-280.
Seonghak Kim. "Ontology-Guided Data Engine Framework for Defense Domain Data Construction" Journal of The Korea Society of Computer and Information 31, no.10 (2026) : 269-280.
