http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
시공간 정보 집중 모델 및 순환 신경망을 사용한 동영상 자동 묘사 시스템
김준영(Junyeong Kim),유창동(Chang D. Yoo) 대한전자공학회 2016 대한전자공학회 학술대회 Vol.2016 No.6
Automatic video description is a challenging problem using computer vision and natural language processing. Previously, researches on automatic video description only engaged either spatial or temporal information, so mutual information of spatio-temporal domain is neglected. In this paper, in order to regard spatial and temporal information simultaneously, a spatio-temporal attention based automatic video description is proposed.
배치-적응형 공간 변형 네트워크를 이용한 도메인에 무관한 얼굴 특징점 추출
마민욱(Minuk Ma),김준영(Junyeong Kim),유창동(Chang D. Yoo) 대한전자공학회 2019 대한전자공학회 학술대회 Vol.2019 No.6
This paper focuses on developing a few-shot facial landmark detector that can operate on the general domain of face. Conventional methods typically used the huge amount of labeled data to perform human facial landmark detection or transferred the knowledge on the human domain to other domains. However, faces share similar geometric structure across various domains ranging from human, animals, cartoons, or even emojis. We propose domain-agnostic few-shot facial landmark detector (DAFF) that possesses both generalizability and dataefficiency. On top of model-agnostic meta-learning (MAML) framework, DAFF is composed of two components: (1) batch-adaptive spatial transformer network with triangle matching loss and (2) facial landmark regressor. Batch-adaptive spatial transformer network first transforms the input image to have a common structure for each facial domain, and facial landmark regressor localizes the position of both eyes and nose. For evaluation, we repurpose LHI-Animal-Faces dataset by adding facial landmark labels on each image. Experimental results show that DAFF outperforms previous methods significantly.
윤현기(Hyungi Yoon),유일수(Ilsu Yoo),박준영(Junyeong Park),박무룡(Mooryong Park) 한국유체기계학회 2008 유체기계 연구개발 발표회 논문집 Vol.2008 No.-
Multi-stage centrifugal pump is widely used over a variety of industrial field such as power and chemistry plant because of its space efficiency and high exit pressure. The 1<SUP>st</SUP> stage pump performance of the multi -stage centrifugal pump is especially important because it is generally composed of the several pumps which are the same geometry along the same shaft. And return channel is used as a connecting passage between the each pump. In the present study, to find out the loss distribution, hydraulic performance and cavitation characteristic, the 1<SUP>st</SUP> stage pump composed of impeller, diffuser and return channel has been investigated numerically by using the k-ω based SST turbulent model, homogeneous flow model and Rayleigh Plesset cavitation model. The power efficiency, total head raise and HPSHR is 0.88, 129m and 2.84m at the designed point, respectively. The numerical analysis of the cavitation characteristic was studied in case of three different calculation models and it was confirmed that only impeller part is needed to find out the cavitation feature in the 1<SUP>st</SUP> stage pump which includes the return channel.