• 한국소프트웨어감정평가학회 논문지
• /
• 제16권2호
• /
• pp.63-68
• /
• 2020

본 논문에서는 딥러닝 프레임워크 기반의 얼굴인증 시스템에 대하여 제안한다. 제안 시스템은 딥러닝 알고리즘을 활용하여 얼굴영역 검출과 얼굴 특징 추출을 수행하고, 결합베이시안 학습 모델을 이용하여 얼굴인증을 수행한다. 제안 얼굴인증 알고리즘에 대한 성능 평가는 다양한 얼굴 사진들로 구성된 데이터베이스를 이용하여 수행하였으며, 한 명에 대한 얼굴 영상은 2장으로 구성하였다. 또한 얼굴인증 실험은 딥 뉴럴 네트워크를 통한 2048차원의 특징과 그 유사성을 측정하기 위해 결합베이시안 알고리즘을 적용하였으며, 얼굴인증에 실패한 동일오율을 계산함으로써 성능평가를 수행하였다. 실험 결과, 딥러닝 특징과 결합베이시안 알고리즘을 사용한 제안 방법은 1.2%의 동일오율을 보였다.

• 천문학회지
• /
• 제55권4호
• /
• pp.131-138
• /
• 2022

We report the discovery of four quasars with M₁₄₅₀ ≳ -25.0 mag at z ~ 5 and supermassive black hole mass measurement for one of the quasars. They were selected as promising high-redshift quasar candidates via deep learning and Bayesian information criterion, which are expected to be effective in discriminating quasars from the late-type stars and high-redshift galaxies. The candidates were observed by the Double Spectrograph on the Palomar 200-inch Hale Telescope. They show clear Lyα breaks at about 7000-8000 Å, indicating they are quasars at 4.7 < z < 5.6. For HSC J233107-001014, we measure the mass of its supermassive black hole (SMBH) using its C _IV λ1549 emission line. The SMBH mass and Eddington ratio of the quasar are found to be ~10⁸ M_⊙ and ~0.6, respectively. This suggests that this quasar possibly harbors a fast growing SMBH near the Eddington limit despite its faintness (L_Bol < 10⁴⁶ erg s^-1). Our 100% quasar identification rate supports high efficiency of our deep learning and Bayesian information criterion selection method, which can be applied to future surveys to increase high-redshift quasar sample.

• 스마트미디어저널
• /
• 제9권1호
• /
• pp.45-50
• /
• 2020

딥 러닝의 발전에 따라 추천시스템에서 딥 러닝 기반의 인공신경망을 활용한 연구가 활발히 진행되고 있다. 특히, RNN(Recurrent Neural Network)기반의 추천시스템은 데이터의 순차적 특성을 고려하기 때문에 추천시스템에서 좋은 성과를 보여주고 있다. 본 연구는 RNN기반의 알고리즘인 GRU(Gated Recurrent Unit)와 세션 기반 병렬 미니배치(Session Parallel mini-batch)기법을 활용한 여행경로 추천 시스템을 제안한다. 본 연구는 top1과 bpr(Bayesian personalized ranking) 오차함수의 앙상블을 통해 추천 성과를 향상시켰다. 또한, 데이터 내에 순차적인 특성을 고려한 RNN기반 추천 시스템은 여행경로에 내재된 여행지의 의미가 반영된 추천이 이루어진다는 것을 확인되었다.

• 한국융합학회논문지
• /
• 제9권4호
• /
• pp.47-55
• /
• 2018

본 논문에서는 얼굴인식 성능 향상을 위해 얼굴 지역 영역 영상들로 학습된 다중개의 심층 합성곱 신경망(Deep Convolutional Neural Network)으로부터 추출된 심층 지역 특징들(Deep local features)을 가중치를 부여하여 결합하는 방법을 제안한다. 제안 방법에서는 지역 영역 집합으로 학습된 다중개의 심층 합성곱 신경망으로부터 추출된 심층 지역 특징들과 해당 지역 영역의 중요도를 나타내는 가중치들을 결합한 특징표현인 '가중치 결합 심층 지역 특징'을 형성한다. 일반화 얼굴인식 성능을 극대화하기 위해, 검증 데이터 집합(validation set)을 사용하여 지역 영역에 해당하는 가중치들을 계산하고 가중치 집합(weight set)을 형성한다. 가중치 결합 심층 지역 특징은 조인트 베이시안(Joint Bayesian) 유사도 학습방법과 최근접 이웃 분류기(Nearest Neighbor classifier)에 적용되어 테스트 얼굴영상의 신원(identity)을 분류하는데 활용된다. 제안 방법은 얼굴영상의 자세, 표정, 조명 변화에 강인하고 기존 최신 방법들과 비교하여 얼굴인식 성능을 향상시킬 수 있음이 체계적인 실험을 통해 검증되었다.

• 전기전자학회논문지
• /
• 제24권4호
• /
• pp.969-974
• /
• 2020

Research into hyper parameter optimization (HPO) has recently revived with interest in models containing many hyper parameters, such as deep neural networks. In this paper, we introduce the most widely used HPO methods, such as grid search, random search, and Bayesian optimization, and investigate their characteristics through experiments. The MNIST data set is used to compare results in experiments to find the best method that can be used to achieve higher accuracy in a relatively short time simulation. The learning rate and weight decay have been chosen for this experiment because these are the commonly used parameters in this kind of experiment.

• 한국농공학회논문집
• /
• 제64권6호
• /
• pp.35-41
• /
• 2022

Heavy snow damage can be prevented in advance with an appropriate security system. To develop the security system, we developed a model that predicts snow depth after a few hours when the snow depth is observed, and utilized it to calculate a failure probability with various types of greenhouses and observed snow depth data. We compared the Markov chain model and Bayesian long short-term memory models with varying input data. Markov chain model showed the worst performance, and the models that used only past snow depth data outperformed the models that used other weather data with snow depth (temperature, humidity, wind speed). Also, the models that utilized 1-hour past data outperformed the models that utilized 3-hour data and 6-hour data. Finally, the Bayesian LSTM model that uses 1-hour snow depth data was selected to predict snow depth. We compared the selected model and the shifting method, which uses present data as future data without prediction, and the model outperformed the shifting method when predicting data after 11-24 hours.

• Smart Structures and Systems
• /
• 제29권1호
• /
• pp.221-235
• /
• 2022

Manual inspection of steel box girders on long span bridges is time-consuming and labor-intensive. The quality of inspection relies on the subjective judgements of the inspectors. This study proposes an automated approach to detect and segment cracks in high-resolution images. An end-to-end cascaded framework is proposed to first detect the existence of cracks using a deep convolutional neural network (CNN) and then segment the crack using a modified U-Net encoder-decoder architecture. A Naïve Bayes data fusion scheme is proposed to reduce the false positives and false negatives effectively. To generate the binary crack mask, first, the original images are divided into 448 × 448 overlapping image patches where these image patches are classified as cracks versus non-cracks using a deep CNN. Next, a modified U-Net is trained from scratch using only the crack patches for segmentation. A customized loss function that consists of binary cross entropy loss and the Dice loss is introduced to enhance the segmentation performance. Additionally, a Naïve Bayes fusion strategy is employed to integrate the crack score maps from different overlapping crack patches and to decide whether a pixel is crack or not. Comprehensive experiments have demonstrated that the proposed approach achieves an 81.71% mean intersection over union (mIoU) score across 5 different training/test splits, which is 7.29% higher than the baseline reference implemented with the original U-Net.

• 응용통계연구
• /
• 제32권2호
• /
• pp.213-227
• /
• 2019

본 논문에서는 베이즈 신경망을 결합한 종단 간 딥러닝 모형을 한국어 음성인식에 적용하였다. 논문에서는 종단 간 학습 모형으로 연결성 시계열 분류기(connectionist temporal classification), 주의 기제, 그리고 주의 기제에 연결성 시계열 분류기를 결합한 모형을 사용하였으며. 각 모형은 순환신경망(recurrent neural network) 혹은 합성곱신경망(convolutional neural network)을 기반으로 하였다. 추가적으로 디코딩 과정에서 빔 탐색과 유한 상태 오토마타를 활용하여 자모음 순서를 조정한 최적의 문자열을 도출하였다. 또한 베이즈 신경망을 각 종단 간 모형에 적용하여 일반적인 점 추정치와 몬테카를로 추정치를 구하였으며 이를 기존 종단 간 모형의 결괏값과 비교하였다. 최종적으로 본 논문에 제안된 모형 중에 가장 성능이 우수한 모형을 선택하여 현재 상용되고 있는 Application Programming Interface (API)들과 성능을 비교하였다. 우리말샘 온라인 사전 훈련 데이터에 한하여 비교한 결과, 제안된 모형의 word error rate (WER)와 label error rate (LER)는 각각 26.4%와 4.58%로서 76%의 WER와 29.88%의 LER 값을 보인 Google API보다 월등히 개선된 성능을 보였다.

• 한국수자원학회:학술대회논문집
• /
• 한국수자원학회 2023년도 학술발표회
• /
• pp.246-246
• /
• 2023

Multi-purpose dams are operated accounting for both physical and socioeconomic factors. This study aims to evaluate the utility of a deep learning algorithm-based model for three multi-purpose dam operation (Seomjin River dam, Juam dam, and Juam Control dam) in Seomjin River. In this study, the Gated Recurrent Unit (GRU) algorithm is applied to predict hourly water level of the dam reservoirs over 2002-2021. The hyper-parameters are optimized by the Bayesian optimization algorithm to enhance the prediction skill of the GRU model. The GRU models are set by the following cases: single dam input - single dam output (S-S), multi-dam input - single dam output (M-S), and multi-dam input - multi-dam output (M-M). Results show that the S-S cases with the local dam information have the highest accuracy above 0.8 of NSE. Results from the M-S and M-M model cases confirm that upstream dam information can bring important information for downstream dam operation prediction. The S-S models are simulated with altered outflows (-40% to +40%) to generate the simulated water level of the dam reservoir as alternative dam operational scenarios. The alternative S-S model simulations show physically inconsistent results, indicating that our deep learning algorithm-based model is not explainable for multi-purpose dam operation patterns. To better understand this limitation, we further analyze the relationship between observed water level and outflow of each dam. Results show that complexity in outflow-water level relationship causes the limited predictability of the GRU algorithm-based model. This study highlights the importance of socioeconomic factors from hidden multi-purpose dam operation processes on not only physical processes-based modeling but also aritificial intelligence modeling.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제12권1호
• /
• pp.392-412
• /
• 2018

Due to the semantic gap problem across different modalities, automatically retrieval from multimedia information still faces a main challenge. It is desirable to provide an effective joint model to bridge the gap and organize the relationships between them. In this work, we develop a deep image annotation and classification by fusing multi-modal semantic topics (DAC_mmst) model, which has the capacity for finding visual and non-visual topics by jointly modeling the image and loosely related text for deep image annotation while simultaneously learning and predicting the class label. More specifically, DAC_mmst depends on a non-parametric Bayesian model for estimating the best number of visual topics that can perfectly explain the image. To evaluate the effectiveness of our proposed algorithm, we collect a real-world dataset to conduct various experiments. The experimental results show our proposed DAC_mmst performs favorably in perplexity, image annotation and classification accuracy, comparing to several state-of-the-art methods.