• Journal of the Korean Institute of Intelligent Systems
• /
• v.17 no.7
• /
• pp.930-938
• /
• 2007

Captions which appear in images include information that relates to the images. In order to obtain the information carried by captions, the methods for text extraction from images have been developed. However, most existing methods can be applied to captions with fixed height of stroke's width. We propose a method which can be applied to various caption size. Our method is based on connected components. And then the edge pixels are detected and grouped into connected components. We analyze the properties of connected components and build a neural network which discriminates connected components which include captions from ones which do not. Experimental data is collected from broadcast programs such as news, documentaries, and show programs which include various height caption. Experimental result is evaluated by two criteria : recall and precision. Recall is the ratio of the identified captions in all the captions in images and the precision is the ratio of the captions in the objects identified as captions. The experiment shows that the proposed method can efficiently extract captions various in size.

• Smart Structures and Systems
• /
• v.6 no.5_6
• /
• pp.595-618
• /
• 2010

This paper describes the application of a wireless sensor network to a 31 meter-tall medieval tower located in the city of Trento, Italy. The effort is motivated by preservation of the integrity of a set of frescoes decorating the room on the second floor, representing one of most important International Gothic artworks in Europe. The specific application demanded development of customized hardware and software. The wireless module selected as the core platform allows reliable wireless communication at low cost with a long service life. Sensors include accelerometers, deformation gauges, and thermometers. A multi-hop data collection protocol was applied in the software to improve the system's flexibility and scalability. The system has been operating since September 2008, and in recent months the data loss ratio was estimated as less than 0.01%. The data acquired so far are in agreement with the prediction resulting a priori from the 3-dimensional FEM. Based on these data a Bayesian updating procedure is employed to real-time estimate the probability of abnormal condition states. This first period of operation demonstrated the stability and reliability of the system, and its ability to recognize any possible occurrence of abnormal conditions that could jeopardize the integrity of the frescos.

• Nuclear Engineering and Technology
• /
• v.55 no.9
• /
• pp.3423-3440
• /
• 2023

In this work, a multivariate time-series machine learning meta-model is developed to predict the transient response of a typical nuclear power plant (NPP) undergoing a steam generator tube rupture (SGTR). The model employs Recurrent Neural Networks (RNNs), including the Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and a hybrid CNN-LSTM model. To address the uncertainty inherent in such predictions, a Bayesian Neural Network (BNN) was implemented. The models were trained using a database generated by the Best Estimate Plus Uncertainty (BEPU) methodology; coupling the thermal hydraulics code, RELAP5/SCDAP/MOD3.4 to the statistical tool, DAKOTA, to predict the variation in system response under various operational and phenomenological uncertainties. The RNN models successfully captures the underlying characteristics of the data with reasonable accuracy, and the BNN-LSTM approach offers an additional layer of insight into the level of uncertainty associated with the predictions. The results demonstrate that LSTM outperforms GRU, while the hybrid CNN-LSTM model is computationally the most efficient. This study aims to gain a better understanding of the capabilities and limitations of machine learning models in the context of nuclear safety. By expanding the application of ML models to more severe accident scenarios, where operators are under extreme stress and prone to errors, ML models can provide valuable support and act as expert systems to assist in decision-making while minimizing the chances of human error.

• Journal of Korean Tunnelling and Underground Space Association
• /
• v.18 no.5
• /
• pp.453-467
• /
• 2016

Overall risks that can occur in a shield TBM tunnelling are studied in this paper. Both the potential risk events that may occur during tunnel construction and their causes are identified, and the causal relationship between causes and events is obtained in a systematic way. Risk impact analysis is performed for the potential risk events and ways to mitigate the risks are summarized. Literature surveys as well as interviews with experts were made for this purpose. The potential risk events are classified into eight categories: cuttability reduction, collapse of a tunnel face, ground surface settlement and upheaval, spurts of slurry on the ground, incapability of mucking and excavation, and water leakage. The causes of these risks are categorized into three areas: geological, design and construction management factors. Bayesian Networks (BN) were established to systematically assess a causal relationship between causes and events. The risk impact analysis was performed to evaluate a risk response level by adopting an Analytic Hierarchy Process (AHP) with the consideration of the downtime and cost of measures. Based on the result of the risk impact analysis, the risk events are divided into four risk response levels and these levels are verified by comparing with the actual occurrences of risk events. Measures to mitigate the potential risk events during the design and/or construction stages are also proposed. Result of this research will be of the help to the designers and contractors of TBM tunnelling projects in identifying the potential risks and for preparing a systematic risk management through the evaluation of the risk response level and the migration methods in the design and construction stage.

• Journal of KIISE:Software and Applications
• /
• v.30 no.3_4
• /
• pp.326-338
• /
• 2003

The objective of a neural network design and model selection is to construct an optimal network with a good generalization performance. However, training data include noises, and the number of training data is not sufficient, which results in the difference between the true probability distribution and the empirical one. The difference makes the teaming parameters to over-fit only to training data and to deviate from the true distribution of data, which is called the overfitting phenomenon. The overfilled neural network shows good approximations for the training data, but gives bad predictions to untrained new data. As the complexity of the neural network increases, this overfitting phenomenon also becomes more severe. In this paper, by taking statistical viewpoint, we proposed an integrative process for neural network design and model selection method in order to improve generalization performance. At first, by using the natural gradient learning with adaptive regularization, we try to obtain optimal parameters that are not overfilled to training data with fast convergence. By adopting the natural pruning to the obtained optimal parameters, we generate several candidates of network model with different sizes. Finally, we select an optimal model among candidate models based on the Bayesian Information Criteria. Through the computer simulation on benchmark problems, we confirm the generalization and structure optimization performance of the proposed integrative process of teaming and model selection.

• The Journal of the Acoustical Society of Korea
• /
• v.42 no.4
• /
• pp.329-344
• /
• 2023

In this study, we comprehensively analyze the generalization performance of various deep learning-based active sonar target classifiers when applied to small and imbalanced active sonar datasets. To generate the active sonar datasets, we use data from two different oceanic experiments conducted at different times and ocean. Each sample in the active sonar datasets is a time-frequency domain image, which is extracted from audio signal of contact after the detection process. For the comprehensive analysis, we utilize 22 Convolutional Neural Networks (CNN) models. Two datasets are used as train/validation datasets and test datasets, alternatively. To calculate the variance in the output of the target classifiers, the train/validation/test datasets are repeated 10 times. Hyperparameters for training are optimized using Bayesian optimization. The results demonstrate that shallow CNN models show superior robustness and generalization performance compared to most of deep CNN models. The results from this paper can serve as a valuable reference for future research directions in deep learning-based active sonar target classification.

• Journal of Korean Society of Transportation
• /
• v.34 no.3
• /
• pp.278-290
• /
• 2016

Maximum traffic flow rate is an important performance measure of operational status in transport networks, and has been considered as a key parameter for transportation operation since a bottleneck in congestion decreases maximum traffic flow rate. Although previous studies for traffic flow analysis have been widely conducted, a detection method for changes in dynamic traffic flow has been still veiled. This paper explores the dynamic traffic flow detection that can be utilized for various traffic operational strategies. Turning point analysis (TPA), as a statistical method, is applied to detect the changes in traffic flow rate. In TPA, Bayesian approach is employed and vehicle arrival is assumed to follow Poisson distribution. To examine the performance of the TPA method, traffic flow data from Jayuro urban expressway were obtained and applied. We propose a novel methodology to detect turning points of dynamic traffic flow in real time using TPA. The results showed that the turning points identified in real-time detected the changes in traffic flow rate. We expect that the proposed methodology has wide application in traffic operation systems such as ramp-metering and variable lane control.

• Journal of the Korea Society of Computer and Information
• /
• v.21 no.6
• /
• pp.9-19
• /
• 2016

In finance literature, stock liquidity showing how stocks can be cashed out in the market has received rich attentions from both academicians and practitioners. The reasons are plenty. First, it is known that stock liquidity affects significantly asset pricing. Second, macroeconomic announcements influence liquidity in the stock market. Therefore, stock liquidity itself affects investors' decision and managers' decision as well. Though there exist a great deal of literature about stock liquidity in finance literature, it is quite clear that there are no studies attempting to investigate the stock liquidity issue as one of decision making problems. In finance literature, most of stock liquidity studies had dealt with limited views such as how much it influences stock price, which variables are associated with describing the stock liquidity significantly, etc. However, this paper posits that stock liquidity issue may become a serious decision-making problem, and then be handled by using data mining techniques to estimate its future extent with statistical validity. In this sense, we collected financial data set from a number of manufacturing companies listed in KRX (Korea Exchange) during the period of 2010 to 2013. The reason why we selected dataset from 2010 was to avoid the after-shocks of financial crisis that occurred in 2008. We used Fn-GuidPro system to gather total 5,700 financial data set. Stock liquidity measure was computed by the procedures proposed by Amihud (2002) which is known to show best metrics for showing relationship with daily return. We applied five data mining techniques (or classifiers) such as Bayesian network, support vector machine (SVM), decision tree, neural network, and ensemble method. Bayesian networks include GBN (General Bayesian Network), NBN (Naive BN), TAN (Tree Augmented NBN). Decision tree uses CART and C4.5. Regression result was used as a benchmarking performance. Ensemble method uses two types-integration of two classifiers, and three classifiers. Ensemble method is based on voting for the sake of integrating classifiers. Among the single classifiers, CART showed best performance with 48.2%, compared with 37.18% by regression. Among the ensemble methods, the result from integrating TAN, CART, and SVM was best with 49.25%. Through the additional analysis in individual industries, those relatively stabilized industries like electronic appliances, wholesale & retailing, woods, leather-bags-shoes showed better performance over 50%.

• Journal of KIISE:Software and Applications
• /
• v.37 no.10
• /
• pp.779-787
• /
• 2010

Since robots have been used widely recently, research about human-robot communication is in process actively. Typically, natural language processing or gesture generation have been applied to human-robot interaction. However, existing methods for communication among robot and human have their limits in performing only static communication, thus the method for more natural and realistic interaction is required. In this paper, an emotional gesture based dialogue management system is proposed for sophisticated human-robot communication. The proposed system performs communication by using the Bayesian networks and pattern matching, and generates emotional gestures of robots in real-time while the user communicates with the robot. Through emotional gestures robot can communicate the user more efficiently also realistically. We used behavior networks as the gesture generation method to deal with dialogue situations which change dynamically. Finally, we designed a usability test to confirm the usefulness of the proposed system by comparing with the existing dialogue system.

• Journal of KIISE
• /
• v.43 no.7
• /
• pp.795-800
• /
• 2016

In image classification, recent CNNs compete with human performance. However, there are limitations in more general recognition. Herein we deal with indoor images that contain too much information to be directly processed and require information reduction before recognition. To reduce the amount of data processing, typically variational inference or variational Bayesian methods are suggested for object detection. However, these methods suffer from the difficulty of marginalizing over the given space. In this study, we propose an image-text integrated recognition system using active vision based on Spatial Transformer Networks. The system attempts to efficiently sample a partial region of a given image for a given language information. Our experimental results demonstrate a significant improvement over traditional approaches. We also discuss the results of qualitative analysis of sampled images, model characteristics, and its limitations.