http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
지상 무기체계 분류 신뢰성 향상을 위한 설명 가능한 인공지능 모델
배기민,이장형 광운대학교 방위사업연구소 2023 선진국방연구 Vol.6 No.3
본 연구는 감시 및 정찰 애플리케이션에서 지상 무기 시스템의 분류 신뢰성을 향상시키기 위한 신뢰할 수 있는 인공 지능(AI) 모델 개발에 중점을 두었다. 전차, 자주포, 다연장 로켓과 같은 군용 물체에 대한 제한된 데이터 가용성으로 인해 제안된 AI 모델은 전이학습 및 미세 조정 기술을 활용하여 이러한 문제를 극복한다. Kaggle에서 공개적으로 사용 가능한 Military-Vehicles 데이터 세트를 사용하여 35개의 딥 러닝 모델을 종합적으로 평가한 결과 MobileNet이 지상 무기 시스템 분류에 가장 적합한 모델임을 확인하였다. 선택한 MobileNet 모델은 5가지 유형의 지상 무기 시스템으로 구성된 데이터 세트에서 테스트했을 때 평균 F1 점수 92%를 달성하였다. 또한 설명 가능한 AI 기술인 Grad-CAM을 적용하여 제안 모델의 의사 결정 프로세스에 대한 통찰력을 제공하고 그 신뢰성을 검증하였다. 훈련비디오에서 추출한 프레임을 사용한 실제 평가는 전차, 자주포 및 다연장 로켓에 대해 유망한 정확도를 보여주었다. 전반적으로 이 연구는 지상 감시 및 정찰 시스템의 성능을 향상시키기 위한 설명 가능하고 신뢰할 수 있는 AI 모델 개발에 기여하였다. This study focused on the development of a reliable artificial intelligence (AI) model to enhance the classification reliability of ground weapon systems for surveillance and reconnaissance applications. The proposed AI model overcomes the limited data availability of military objects such as tanks, canons, and multiple-launch rockets by leveraging transfer learning and fine-tuning techniques. A comprehensive evaluation of 35 deep learning models using the publicly available Military-Vehicles dataset on Kaggle identified MobileNet as the most suitable model for ground weapon system classification. The selected MobileNet model achieved an average F1 score of 92% when tested on a dataset comprising five types of ground-weapon systems. In addition, the application of the explainable AI technique Grad-CAM provided insights into the decision-making process of the proposed model and verified its reliability. Real-world evaluations using frames extracted from training videos demonstrated promising accuracy for tanks, canons, and multiple-launch rockets. However, challenges related to object occlusion and the absence of target objects in the images were observed, which resulted in misclassifications. Overall, this study contributes to the development of explainable and reliable AI models for enhancing the performance of ground surveillance and reconnaissance systems.
퓨샷 학습과 삼중항 손실을 이용한 지상 무기체계 분류 모델의 확장성 연구
배기민,이장형 육군사관학교 화랑대연구소 2023 한국군사학논집 Vol.79 No.3
This paper proposes an extensible ground weapon system classification model based on a dataset with 11 categories obtained through web crawling and publicly available operation and training videos. A base model is trained using this dataset to extract discriminative features, similar to conventional supervised learning. By applying few-shot learning and triplet loss, an extended model is trained based on the previous training. The evaluation shows the base model achieved 93.35% to 97.3% accuracy for 5 categories and 93.54% to 99% accuracy for 11 categories. The extended model, achieved 72.18% to 75.54% accuracy when trained on 5 categories and tested on 11. These experiments demonstrate the effectiveness of few shot learning and triplet loss in creating an extended model capable of predicting new, untrained categories even with limited data.
배기민,김강웅,이상민 한국수산과학회 2015 Fisheries and Aquatic Sciences Vol.18 No.2
We evaluated the effects of rice distillers dried grain (DDG) as a partial replacement for fish meal in the practical diet on growth performance, feed utilization, and body composition of juvenile olive flounder Paralichthys olivaceus. Six isonitrogenous and isocaloric diets were formulated to contain 0%, 7%, 14%, 21%, 28%, and 35% DDG (designated DDG0, DDG7, DDG14, DDG21, DDG28, and DDG35, respectively). Three replicate groups of juvenile olive flounder averaging 9.6 ± 0.2 g were fed one of the experimental diets to visual satiety twice daily for 8 weeks. Neither survival nor daily feed intake was affected by the dietary DDG levels. Weight gain of the flounder fed the DDG28 and DDG35 diets was lower than that of flounder fed the DDG7 diet. The feed efficiency of flounder fed the DDG28 diet was lower than that of flounder fed the DDG0, DDG7, and DDG14 diets. The protein efficiency ratio of flounder fed the DDG28 diet was lower than that of flounder fed the DDG7 diet. The proximate composition of muscle was not affected by the dietary DDG levels. The plasma contents of total protein, glucose, cholesterol, glutamate oxaloacetate transaminase, phospholipid, and triglyceride were not affected by the dietary DDG levels. The results of this experiment suggest that DDG has the potential to replace fish meal and could be used up to 21% DDG without any negative effects on the growth and feed utilization of juvenile flounder.