• 대한전자공학회논문지SP
• /
• 제44권1호
• /
• pp.40-48
• /
• 2007

이 논문은 다중 스케일 베이지안 관점에서 다층 퍼셉트론과 마코프 랜덤 필드를 사용한 새로운 결 분할 방법을 제안한다. 다층 퍼셉트론의 출력은 사후 확률을 모델링하므로 본 논문에서는 다중 스케일 웨이블릿 계수들을 다층 퍼셉트론의 입력으로 사용한다. 다층 퍼셉트론으로부터 구한 사후 확률과 MAP (maximum a posterior) 분류를 이용하여 각 스케일에서 결 분류를 수행한다. 또한 가장 섬세한 스케일에서 더 개선된 분할 결과를 얻기 위하여 모든 스케일에서 MAP 분류 결과들을 거친 스케일에서 섬세한 스케일까지 차례로 융합한다. 이런 과정은 한 스케일에서의 분류 정보와 그 인접한 보다 거친 스케일에서 얻어지는 문맥과 관련한 연역적 정보를 이용하여 MAP 분류를 행함으로써 이루어진다. 이 융합 과정에서, MRF (Markov random fields) 사전 모델이 평탄화 제한자로서 동작하고, 깁스 샘플러 (Gibbs sampler)는 MAP 분류기로서 동작한다. 제안한 분할 방법은 HMT (Hidden Markov Trees) 모델과 HMTseg 알고리즘을 이용한 결 분할 방법보다 더 좋은 성능을 보인다.

• Structural Engineering and Mechanics
• /
• 제54권2호
• /
• pp.337-362
• /
• 2015

This paper presents a feasibility study on structural damage alarming and localization of long-span cable-supported bridges using multi-novelty indices formulated by monitoring-derived modal parameters. The proposed method which requires neither structural model nor damage model is applicable to structures of arbitrary complexity. With the intention to enhance the tolerance to measurement noise/uncertainty and the sensitivity to structural damage, an improved novelty index is formulated in terms of auto-associative neural networks (ANNs) where the output vector is designated to differ from the input vector while the training of the ANNs needs only the measured modal properties of the intact structure under in-service conditions. After validating the enhanced capability of the improved novelty index for structural damage alarming over the commonly configured novelty index, the performance of the improved novelty index for damage occurrence detection of large-scale bridges is examined through numerical simulation studies of the suspension Tsing Ma Bridge (TMB) and the cable-stayed Ting Kau Bridge (TKB) incurred with different types of structural damage. Then the improved novelty index is extended to formulate multi-novelty indices in terms of the measured modal frequencies and incomplete modeshape components for damage region identification. The capability of the formulated multi-novelty indices for damage region identification is also examined through numerical simulations of the TMB and TKB.

• 한국인공지능학회지
• /
• 제11권3호
• /
• pp.17-22
• /
• 2023

In this paper, we propose a black ice detection system using Convolutional Neural Networks (CNNs). Black ice poses a serious threat to road safety, particularly during winter conditions. To overcome this problem, we introduce a CNN-based architecture for real-time black ice detection with an encoder-decoder network, specifically designed for real-time black ice detection using thermal images. To train the network, we establish a specialized experimental platform to capture thermal images of various black ice formations on diverse road surfaces, including cement and asphalt. This enables us to curate a comprehensive dataset of thermal road black ice images for a training and evaluation purpose. Additionally, in order to enhance the accuracy of black ice detection, we propose a multi-scale dilation convolution feature fusion (MsDC-FF) technique. This proposed technique dynamically adjusts the dilation ratios based on the input image's resolution, improving the network's ability to capture fine-grained details. Experimental results demonstrate the superior performance of our proposed network model compared to conventional image segmentation models. Our model achieved an mIoU of 95.93%, while LinkNet achieved an mIoU of 95.39%. Therefore, it is concluded that the proposed model in this paper could offer a promising solution for real-time black ice detection, thereby enhancing road safety during winter conditions.

• 한국방송∙미디어공학회:학술대회논문집
• /
• 한국방송∙미디어공학회 2021년도 하계학술대회
• /
• pp.307-310
• /
• 2021

X-ray 영상 검사·검출 문제에 기존 딥러닝 모델을 사용하려는 시도들이 존재해왔고, 합성곱 신경망의 강력한 표현력 덕분에 대체로 준수한 성능이 보장되었다. 그러나 문제의 특성에 따라 기대한 만큼의 분류 및 검출 성능이 나오지 않는 경우가 존재한다. 이는 1) 검출 대상의 스케일이 다양하거나, 2) X-ray 영상은 흑백 영상으로 미세한 특징을 학습하기 어렵거나, 3) 지도학습을 하기에는 학습 데이터의 양이 부족하기 때문인 것이 주요 원인들이다. 본 논문에서는 다양한 스케일의 특징맵을 추출하여 종합적으로 학습하는 신경망을 통해, '생선살 X-ray 영상' 데이터셋에서 '생선 가시' 이물질 class가 모델 내에서 어떻게 학습되는지를 살펴본다. 그리고 X-ray 영상의 경우, 이물질 class를 크기별로 새롭게 labeling하여 성능 개선이 일어날 수 있음을 보인다. 또한 Multi-scale CAM을 통해 class에 따른 활성화 정도를 시각화하여 모델을 직관적으로 분석할 수 있음을 보일 것이다.

• 천문학회보
• /
• 제44권2호
• /
• pp.53.3-53.3
• /
• 2019

Recent work by Ravanbakhsh et al. (2017), Mathuriya et al. (2018) showed that convolutional neural networks (CNN) can be trained to predict cosmological parameters from the visual shape of the large scale structure, i.e. the filaments, clusters and voids of the cosmic density field. These preliminary works used the dark matter density field at redshift zero. We build upon these works by considering realistic mock galaxy catalogues that mimic true observations. We construct light-cones that span the redshift range appropriate for current and near future cosmological surveys such as LSST, EUCLID, WFIRST etc. In summary, we propose a novel multi-image input CNN to track the evolution in the morphology of large scale structures over cosmic time to constrain cosmology and the expansion history of the Universe.

• 로봇학회논문지
• /
• 제12권3호
• /
• pp.313-321
• /
• 2017

One of the most frequently performed tasks in human-robot interaction (HRI), intelligent vehicles, and security systems is face related applications such as face recognition, facial expression recognition, driver state monitoring, and gaze estimation. In these applications, accurate head pose estimation is an important issue. However, conventional methods have been lacking in accuracy, robustness or processing speed in practical use. In this paper, we propose a novel method for estimating head pose with a monocular camera. The proposed algorithm is based on a deep neural network for multi-task learning using a small grayscale image. This network jointly detects multi-view faces and estimates head pose in hard environmental conditions such as illumination change and large pose change. The proposed framework quantitatively and qualitatively outperforms the state-of-the-art method with an average head pose mean error of less than $4.5^{\circ}$ in real-time.

• Journal of Information Processing Systems
• /
• 제17권2호
• /
• pp.411-425
• /
• 2021

In this paper, a vehicle recognition algorithm based on deep convolutional neural network and compression dictionary is proposed. Firstly, the network structure of fine vehicle recognition based on convolutional neural network is introduced. Then, a vehicle recognition system based on multi-scale pyramid convolutional neural network is constructed. The contribution of different networks to the recognition results is adjusted by the adaptive fusion method that adjusts the network according to the recognition accuracy of a single network. The proportion of output in the network output of the entire multiscale network. Then, the compressed dictionary learning and the data dimension reduction are carried out using the effective block structure method combined with very sparse random projection matrix, which solves the computational complexity caused by high-dimensional features and shortens the dictionary learning time. Finally, the sparse representation classification method is used to realize vehicle type recognition. The experimental results show that the detection effect of the proposed algorithm is stable in sunny, cloudy and rainy weather, and it has strong adaptability to typical application scenarios such as occlusion and blurring, with an average recognition rate of more than 95%.

• 한국정보처리학회논문지
• /
• 제4권7호
• /
• pp.1881-1892
• /
• 1997

본 논문에서는 확률적 VQ 네트워크와 계층적 구조를 가지는 다단계 인식기를 이용한 인쇄체 한자 인식 방법을 제안한다. 대용량 신경망은 구현하기가 매우 어렵기 때문에 모듈화된 신경망을 이용하였으며, 이 과정에서 발생되는 문제점을 확률적 신경망 모델을 이용으로 제거하였다. 또한 엔트로피 이론을 적용하여 오인식률이 높은 혼동 문자쌍에 대하여 재분류를 수행하였다. 실험대상은 KSC5601 코드의 한자 4,888자 중, 동자이음문자를 제외한 4,619자로 하였으며, 학습 데이타와 실험 데이타에 대하여 실험결과, 각각 평균 99.33%, 92.83%의 인식률과 초당 4-5자의 인식속도를 얻음으로써 본 방법의 유효성을 보였다.

• 한국정보처리학회:학술대회논문집
• /
• 한국정보처리학회 2019년도 추계학술발표대회
• /
• pp.1004-1006
• /
• 2019

Different scale of information is an important component in computer vision systems. Recently, there are considerable researches on utilizing multi-scale information to solve the scale-invariant problems, such as GoogLeNet and FPN. In this paper, we introduce the notion of different scale information fusion (DSIF) and show that it has a significant effect on the performance of object recognition systems. We analyze the DSIF in several architecture designs, and the effect of nonlinear activations, dropout, sub-sampling and skip connections on it. This leads to clear suggestions for ways of the DSIF to choose.

• Biotechnology and Bioprocess Engineering:BBE
• /
• 제5권5호
• /
• pp.355-361
• /
• 2000

A recycling reactor system operated under sequential anoxic and oxic conditions for the treatment of swine wastewater has been developed, in which piggery slurry is fermentatively and aerobically treated and then part of the effluent is recycled to the pigsty. This system significantly removes offensive smells (at both the pigsty and the treatment plant), BOD and others, and may be cost effective for small-scale farms. The most dominant heterotrophic were, in order, Alcaligenes faecalis, Brevundimonas diminuta and Streptococcus sp., while lactic acid bacteria were dominantly observed in the anoxic tank. We propose a novel monitoring system for a recycling piggery slurry treatment system through the use of neural networks. In this study, we tried to model the treatment process for each tank in the system (influent, fermentation, aeration, first sedimentation and fourth sedimentation tanks) based upon the population densities of the heterotrophic and lactic acid bacteria. Principal component analysis(PCA) was first applied to identify a relationship between input and output. The input would be microbial densities and the treatment parameters, such as population densities of heterotrophic and lactic acid bacteria, suspended solids(SS), COD, NH$_4$(sup)+-N, ortho-phosphorus (o-P), and total-phosphorus (T-P). then multi-layer neural networks were employed to model the treatment process for each tank. PCA filtration of the input data as microbial densities was found to facilitate the modeling procedure for the system monitoring even with a relatively lower number of imput. Neural network independently trained for each treatment tank and their subsequent combined data analysis allowed a successful prediction of the treatment system for at least two days.