http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
기동무기체계 분류를 위한 전이학습 기반 앙상블 모델 연구
황진용(Jinyong Hwang),최보아(BoAh Choi),이종혁(JongHyuk Lee),배지훈(Ji-Hoon Bae) 한국정보기술학회 2021 한국정보기술학회논문지 Vol.19 No.12
Neural network-based deep learning technology is most widely used in image detection, classification, and identification. However, because a large of image data is required, its use may be limited in specific areas such as military fields, where data acquisition is limited. Therefore, research on how to achieve higher performance with a small amount of data is in progress. In this paper, after collecting images of tanks and armored vehicles that are maneuver weapons systems, an ensemble was performed to obtain the average of the predicted values for each model derived through transfer learning. As a result of analyzing the accuracy of image classification for various ensemble combinations of multiple models, for the image dataset considered in this paper, it was confirmed that 98.5% classification accuracy could be achieved with only a combination of two heterogeneous models, InceptionV3 and DenseNet121 ensemble.
전이학습 기반 불균형 앙상블 모델의 가중치 조정에 따른 성능 분석
황진용(Jinyong Hwang),배지훈(Ji-Hoon Bae) 한국정보기술학회 2021 Proceedings of KIIT Conference Vol.2021 No.11
전이학습은 적은 데이터를 이용하더라도 과대적합을 쉽게 극복하면서 정확도를 높일 수 있는 방법으로 현재까지 널리 활용되고 있으며, 여기에 두 가지 이상 모델들을 앙상블하는 기법은 이보다 더 좋은 결과를 제공할 수 있다. 다양한 사전 학습 완료된 모델들을 사용하는 전이학습의 경우 각 모델 별로 정확도에 차이가 있고, 하이퍼파라미터 설정에 따라 성능 차이가 발생할 수 있다. 따라서, 앙상블 수행 시 주로 예측을 평균하기 때문에 어느 정도 유사한 성능을 가지는 모델들의 경우 앙상블 효과가 높아지지만, 이와 반대로 성능의 불균형이 발생될 경우 앙상블 효과가 오히려 저하될 수 있다. 이에 본 논문에서는 정확도 불균형이 있는 이종모델들에 대하여 앙상블 가중치를 조정하였을 때 앙상블 정확도 성능을 분석해본다. Transfer learning is widely used to date as a method that can easily overcome overfitting and increase accuracy even with a small amount of data, and a technique that ensembles two or more models can provide better results than this. In the case of transfer learning using various pre-trained model, there are differences in accuracy for each model, and performance differences can occur depending on hyper-parameter settings. Therefore, since the prediction is mainly averaged for performing an ensemble, the ensemble effect is increased for models when the models have similar accuracies. However, if an imbalance in performance occurs, the ensemble effect can be decreased. This paper analyzes the ensemble accuracy performance when the ensemble weights are adjusted for heterogeneous models with accuracy imbalance.
연규석 ( Kyuseok Yeon ),황진용 ( Jinyong Hwang ),김관규 ( Kwankyu Kim ),연정흠 ( Jungheum Yeon ) 한국농공학회 2014 한국농공학회 학술대회초록집 Vol.2014 No.-
이 연구는 노후 된 수로교의 리모델링 공법을 개발하고 이 기술을 실용화하기 위한 연구로서, 얻어진 결과를 요약하면 다음과 같다. 1. 프리캐스트 폴리머 콘크리트 제품을 이용한 교체 공법을 현장에 시험 시공한 바 만족할만한 결과를 얻을 수 있었다. 2. 기존의 보수 보강 기술에 비해 공사기간을 단축 할 수 있고, 시공성을 향상시킬 수 있기 때문에 충분한 경쟁력이 있음을 확인 할 수 있었다. 3. 폴리머 콘크리트 수로교의 신축량의 이론값을 실측치와 비교하여 본 바 그 값이 거의 일치함을 확인 할 수 있었다. 4. 프리캐스트 폴리머 콘크리트 제품으로 리모델링하는 공법은 노후 된 개거부를 철거하고 철근콘크리트로 재시공 하는 공법에 비해 경제성 측면에서 매우 유리한 것으로 나타났다. This investigation was conducted to develop remodeling methods for decrepit aqueducts and put the resulting technology to practical use. Base on the results from this investigation, then following conclusions are drawn: 1. Demonstrative construction was made of the replacement method using precast polymer concrete products in the fields, showing satisfactory results. 2. Construction period could be shortened as compared to rehabilitation method currently available and constructability could be improved, thus proving the technology's sufficient competitiveness. 3. Theoretical values of expansion length of polymer concrete aqueducts proved almost identical to measured values thereof. 4. Remodeling method with precast polymer concrete products was showing great economic merit compare to rebuilding with reinforced concretes.
지상무기체계 성능개량을 위한 개방형 시스템 아키텍처 및 데이터 모델 설계
한동휘(Donghwi Han),황진용(Jinyong Hwang),배지훈(Ji-Hoon Bae),이종혁(JongHyuk Lee) 한국정보기술학회 2021 한국정보기술학회논문지 Vol.19 No.12
Weapon system improvement is an efficient acquisition method that can shorten the deployment time and achieve maximum effect with minimum cost. In the case of a weapon system to which an open system architecture is applied during the improvement process, it has the advantage of relatively easy replacement and addition of major parts by utilizing its structural advantages. In this paper, the open system architecture was applied to the ground weapon system in operation after the improvement was completed. In addition, the data distribution service, a data exchange method, among the open system architecture was applied. Based on this, a data model for replacing and integrating the radar of the ground weapon system with the latest sensor was designed according to the characteristics of the weapon system. As a result, it was possible to predict the cost and time savings of weapon system modification, and the system applicability of the designed data model was confirmed through the verification test.