전장 이미지 기반 YOLO-MobileNet 전차 전투피해평가 연구 = A Study on Tank Battle Damage Assessment Using YOLO?MobileNet Based on Battlefield Imagery

다국어 초록 (Multilingual Abstract)

A Study on Tank Battle Damage Assessment Using YOLO–MobileNet Based on Battlefield Imagery Moon, Ji Man(성명) Weapon System(전공명) Graduate School of Defense Management Korea National Defense University In modern warfare, Battle Damage Assessment (BDA) plays a critical role in maintaining combat power and ensuring operational continuity. To overcome the limitations of manual image interpretation, this study proposes a YOLO–MobileNet hybrid framework for automatic tank damage recognition using real battlefield imagery from the Russia–Ukraine conflict. The proposed system integrates a YOLOv8 detector for high-precision tank localization and a MobileNetV2 classifier for simultaneous prediction of damage type and severity. A META labeling scheme, designed based on U.S. Army maintenance criteria (AR 750-1), encodes both attributes into a unified structure, enabling the output to directly support tactical and maintenance decision-making. Experimental results demonstrated strong performance, achieving a detection accuracy (mAP@0.5) of 0.921, damage-type classification accuracy of 71.5% (F1-score 0.67), damage-severity accuracy of 64.9% (F1-score 0.65), and a real-time processing speed of 118 FPS. These results indicate the framework’s high feasibility for real-time UAV-based battlefield analysis, surpassing prior simulation-based studies in both realism and computational efficiency. This study validates the operational applicability of automated BDA for real-time decision support in combat environments. Future work will focus on refining the META labeling scheme and integrating Transformer-based architectures with multi-sensor fusion to develop a more precise and scalable BDA framework. Key words: Battle Damage Assessment, Battlefield Imagery, YOLO, MobileNet, META Labeling

목차 (Table of Contents)

• 제1장 서론 1
• 제2장 관련 연구 5
• 제1절 전투피해평가(BDA) 자동화 연구5
• 제2절 전장 이미지 탐지/분류 연구 10
• 제3절 정비 기준 손상 평가 체계 17
• 제3장 연구 방법 19
• 제1절 YOLO-MobileNet 하이브리드 구조 설계 19
• 제2절 META 라벨링 체계 22
• 제3절 데이터셋 구성 및 증강 전략 24
• 제4절 평가 지표 설정 27
• 제4장 실험 및 결과 31
• 제1절 실험 환경 및 설계 31
• 제2절 실험 결과 34
• 제3절 결과 분석 38
• 제5장 결론41
• 참고문헌44
• 영문요약46