• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.348-353
• /
• 2022

최근 신경망 기반의 언어모델이 발전함에 따라 대부분의 검색 모델에서는 Bi-encoder를 기반으로한 Dense retrieval 모델에 대한 연구가 진행되고 있다. 특히 DPR은 BM25를 통해 정답 문서와 유사한 정보를 가진 하드 네거티브를 사용하여 대조학습을 통해 성능을 더욱 끌어올린다. 그러나 BM25로 검색된 하드 네거티브는 term-base의 유사도를 통해 뽑히기 때문에, 의미적으로 비슷한 내용을 갖는 하드 네거티브의 역할을 제대로 수행하지 못하고 대조학습의 효율성을 낮출 가능성이 있다. 따라서 DRP의 대조학습에서 하드 네거티브의 역할을 본질적으로 수행할 수 있는 문서를 샘플링 하는 방법을 제시하고, 이때 얻은 하드 네거티브의 집합을 주기적으로 업데이트 하여 효과적으로 대조학습을 진행하는 방법을 제안한다. 지식 기반 대화 데이터셋인 MultiDoc2Dial을 통해 평가를 수행하였으며, 실험 결과 기존 방식보다 더 높은 성능을 나타낸다.

• Proceedings of the IEEK Conference
• /
• 2001.09a
• /
• pp.297-300
• /
• 2001

본 논문에서는 동영상의 계층적 부호화의 효율을 높이기 위한 방안에 대해 연구하였다. 단일 계층부호화에 비해 다 계층부호화는 계산량이 많아진다. 따라서 계층적 부호화의 장점을 살리고 단점을 보완하는 방안을 제시하였다. 우선 인코더에서 고급계층의 복잡도를 줄이기 위하여 고급계층의 참조 형태를 P-VOP (Prediction-Video Object Plane)만으로 정한다. 고급계층의 참조 영역으로 사용되는 업샘플링된 VOP의 횟수를 줄여서 업샘플링에 따른 계산량을 줄인다. 그리고 고급계층의 비트율을 조절하여 Traffic shaping 효과도 얻을 수 있다. 이러한 방법들을 통해 단일 계층 부호화에 비해 다 계층부호화의 장점을 살리고 단점을 보완하는 코덱을 제안한다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2015.07a
• /
• pp.73-74
• /
• 2015

3 차원 컨텐츠 제작은 많은 관심을 받고 있는 분야이다. 변위 지도로 표현 가능한 깊이 정보는 3 차원 컨텐츠를 생성하는데 필수적이다. 본 논문에서는 깊이 카메라 및 스테레오 카메라를 이용하여 정확한 변위 지도를 생성하는 방법을 제안한다. 제안하는 방법은 스테레오 영상 사이의 변위를 예측하기 위해서 깊이 카메라 정보를 3 차원 워핑 방식에 의해서 좌우 카메라 위치로 투영한다. 투영된 깊이 정보는 스테레오 영상의 크기에 맞춰서 업샘플링된다. 최종적으로 업샘플링된 깊이 카메라 정보와 스테레오 정보가 결합되어 정확한 변위 지도를 생성한다. 실험 결과는 제안하는 방법이 기존의 단일 센서를 이용한 방식에 비해서 좀더 정확한 결과를 생성함을 보여준다.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.29 no.5
• /
• pp.471-478
• /
• 2011

High resolution Pan-sharpening technique is becoming increasingly important in the field of remote sensing image analysis as an essential image processing to improve the spatial resolution of original multispectral image. The general scheme of pan-sharpening technique consists of upsampling process of multispectral image and high-pass detail injection process using the panchromatic image. The upsampling process, however, brings about image blurring, and this lead to spectral distortion in the pan-sharpening process. In order to solve this problem, this paper presents a new method that adopts image restoration techniques based on optimization theory in the pan-sharpening process, and evaluates its efficiency and application possibility. In order to evaluate the effect of image restoration techniques on the pansharpening process, the result obtained using the existing method that used bicubic interpolation were compared visually and quantitatively with the results obtained using image restoration techniques. The quantitative comparison was done using some spectral distortion measures for use to evaluate the quality of pan-sharpened image.

• Journal of the Korea Computer Graphics Society
• /
• v.18 no.3
• /
• pp.17-27
• /
• 2012

In this paper, we describe a case study of the film 'Sector 7' which was produced by technologies applied fluid simulation. For the CG scenes in the movie which include highly detailed fluid motions, we used smoothed particle hydrodynamics(SPH) technique to express subtle movements of seawater from a crashed huge tank, and used hybrid simulation method of particles and levelsets to describe bursting water from a submarine's broken canopy. We also used detonation shock dynamics(DSD) technique for detailed flame simulations to produce a burning monster, the film"s main character. At this point, the divergence-free vortex particle method was applied to conserve the incompressible property of fluids. In addition, we used an upsampling method to achieve more efficient video production. Consequently, we could produce the high-quality visual effects by using the domestic technologies.

• Journal of Broadcast Engineering
• /
• v.26 no.3
• /
• pp.283-295
• /
• 2021

This paper proposes a method for generating dense depth map using information of color images and depth map generated based on lidar based on self-organizing map. The proposed depth map upsampling method consists of an initial depth prediction step for an area that has not been acquired from LiDAR and an initial depth filtering step. In the initial depth prediction step, stereo matching is performed on two color images to predict an initial depth value. In the depth map filtering step, in order to reduce the error of the predicted initial depth value, a self-organizing map technique is performed on the predicted depth pixel by using the measured depth pixel around the predicted depth pixel. In the process of self-organization map, a weight is determined according to a difference between a distance between a predicted depth pixel and an measured depth pixel and a color value corresponding to each pixel. In this paper, we compared the proposed method with the bilateral filter and k-nearest neighbor widely used as a depth map upsampling method for performance comparison. Compared to the bilateral filter and the k-nearest neighbor, the proposed method reduced by about 6.4% and 8.6% in terms of MAE, and about 10.8% and 14.3% in terms of RMSE.

• Journal of the Korea institute for structural maintenance and inspection
• /
• v.14 no.5
• /
• pp.119-127
• /
• 2010

This paper provides a new approach for durability prediction of reinforced concrete structures exposed to carbonation. In this method, the prediction can be updated successively by a Bayes' theorem when additional data are available. The stochastic properties of model parameters are explicitly taken into account in the model. To simplify the procedure of the model, the probability of the durability limit is determined based on the samples obtained from the Latin Hypercube Sampling(LHS) technique. The new method may be very useful in design of important concrete structures and help to predict the remaining service life of existing concrete structures which have been monitored. For using the new method, in which the prior distribution is developed to represent the uncertainties of the carbonation velocity using data of concrete structures(3700 specimens) in Korea and the likelihood function is used to monitor in-situ data. The posterior distribution is obtained by combining a prior distribution and a likelihood function. Efficiency of the LHS technique for simulation was confirmed through a comparison between the LHS and the Monte Calro Simulation(MCS) technique.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.06a
• /
• pp.82-83
• /
• 2019

본 논문에서는 휘도 및 색차 성분에 따른 SHVC 계층간 참조 픽처 생성 방법의 부호화 성능을 분석한다. SHVC 상위 계층에서는 하위 계층의 픽처를 DCT-IF 기반 업샘플링하여 사용한다. 상위 계층의 부호화 성능을 높이기 위해 딥러닝 기반 필터링을 이용하여 휘도, 색차 성분의 고주파 신호 복원이 부호화 성능에 미치는 영향을 분석한다. 기존 Y 성분에만 VDSR 네트워크를 이용하여 필터링을 적용하였을 때보다 색차 성분까지 필터링을 진행할 경우 최대 2.18%, 평균 1.5% 감소된 결과를 보였다.

• Journal of the Korean Society for Aeronautical & Space Sciences
• /
• v.42 no.1
• /
• pp.82-89
• /
• 2014

This paper discusses the methods for reducing the sampling time quantization errors of the body dither type ring laser gyroscope. A ring laser gyroscope has the angle quantization error which is generated by the frequency counting method of the laser beat signal and sampling time quantization error which is generated by the demodulation method for eliminating the body dithering in which the sampling periods are fitted to the dither periods. Generally, because the dither periods are longer than the calculation periods of the inertial navigation system, vehicle navigation errors are produced by long time attitude update missing during the vehicle move with a high dynamical motion. In this paper, the double demodulation method is proposed for reducing the sampling time quantization error and its effects under the dynamic situation are confirmed by simulation.

• The Journal of Bigdata
• /
• v.2 no.2
• /
• pp.95-104
• /
• 2017

In 2014, Seoul Metropolitan Government launched a response service aimed at responding promptly to civil complaints. The complaints received are categorized based on their content and sent to the department in charge. If this part can be automated, the time and labor costs will be reduced. In this study, we collected 17,700 cases of complaints for 7 years from June 1, 2010 to May 31, 2017. We compared the XGBoost with RandomForest and confirmed the suitability of Korean text classification. As a result, the accuracy of XGBoost compared to RandomForest is generally high. The accuracy of RandomForest was unstable after upsampling and downsampling using the same sample, while XGBoost showed stable overall accuracy.