• Korean Journal of Cognitive Science
• /
• v.10 no.3
• /
• pp.1-15
• /
• 1999

Based on known physiological and psychophysical results, a neural network model for visual selection, called FeaureGate is proposed. The model consists of a hierarchy of spatial maps. and the flow of information from each level of the hierarchy to the next is controlled by attentional gates. The gates are jointly controlled by a bottom-up system favoring locations with unique features. and a top-down mechanism favoring locations with features designated as target features. The present study focuses on the top-down mechanism of the FeatureGate model that produces results similar to Moran and Desimone's (1985), which many current models have failed to explain, The FeatureGate model allows a consistent interpretation of many different experimental results in visual attention. including parallel feature searches and serial conjunction searches. attentional gradients triggered by cuing, feature-driven spatial selection, split a attention, inhibition of distractor locations, and flanking inhibition. This framework can be extended to produce a model of shape recognition using upper-level units that respond to configurations of features.

• Journal of Korea Water Resources Association
• /
• v.45 no.7
• /
• pp.631-641
• /
• 2012

In this study, satellite data (MTSAT-1R), a numerical weather prediction model, RDAPS (Regional Data Assimilation and Prediction System) output, ground weather station data, and artificial neural networks were used to improve the accuracy of summer rainfall forecasts. The developed model was applied to the Seoul station to forecast the rainfall at 3, 6, 9, and 12-hour lead times. Also to reflect the different weather conditions during the summer season which is related to the frontal precipitation and the cyclonic precipitation such as Jangma and Typhoon, the neural network models were formed for two different periods of June-July and August-September respectively. The rainfall forecast model was trained during the summer season of 2006 and 2008 and was verified for that of 2009 based on the data availability. The results demonstrated that the model allows us to get the improved rainfall forecasts until lead time of 6 hour, but there is still a large room to improve the rainfall forecast skill.

• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.163-177
• /
• 2019

As smartphones are getting widely used, human activity recognition (HAR) tasks for recognizing personal activities of smartphone users with multimodal data have been actively studied recently. The research area is expanding from the recognition of the simple body movement of an individual user to the recognition of low-level behavior and high-level behavior. However, HAR tasks for recognizing interaction behavior with other people, such as whether the user is accompanying or communicating with someone else, have gotten less attention so far. And previous research for recognizing interaction behavior has usually depended on audio, Bluetooth, and Wi-Fi sensors, which are vulnerable to privacy issues and require much time to collect enough data. Whereas physical sensors including accelerometer, magnetic field and gyroscope sensors are less vulnerable to privacy issues and can collect a large amount of data within a short time. In this paper, a method for detecting accompanying status based on deep learning model by only using multimodal physical sensor data, such as an accelerometer, magnetic field and gyroscope, was proposed. The accompanying status was defined as a redefinition of a part of the user interaction behavior, including whether the user is accompanying with an acquaintance at a close distance and the user is actively communicating with the acquaintance. A framework based on convolutional neural networks (CNN) and long short-term memory (LSTM) recurrent networks for classifying accompanying and conversation was proposed. First, a data preprocessing method which consists of time synchronization of multimodal data from different physical sensors, data normalization and sequence data generation was introduced. We applied the nearest interpolation to synchronize the time of collected data from different sensors. Normalization was performed for each x, y, z axis value of the sensor data, and the sequence data was generated according to the sliding window method. Then, the sequence data became the input for CNN, where feature maps representing local dependencies of the original sequence are extracted. The CNN consisted of 3 convolutional layers and did not have a pooling layer to maintain the temporal information of the sequence data. Next, LSTM recurrent networks received the feature maps, learned long-term dependencies from them and extracted features. The LSTM recurrent networks consisted of two layers, each with 128 cells. Finally, the extracted features were used for classification by softmax classifier. The loss function of the model was cross entropy function and the weights of the model were randomly initialized on a normal distribution with an average of 0 and a standard deviation of 0.1. The model was trained using adaptive moment estimation (ADAM) optimization algorithm and the mini batch size was set to 128. We applied dropout to input values of the LSTM recurrent networks to prevent overfitting. The initial learning rate was set to 0.001, and it decreased exponentially by 0.99 at the end of each epoch training. An Android smartphone application was developed and released to collect data. We collected smartphone data for a total of 18 subjects. Using the data, the model classified accompanying and conversation by 98.74% and 98.83% accuracy each. Both the F1 score and accuracy of the model were higher than the F1 score and accuracy of the majority vote classifier, support vector machine, and deep recurrent neural network. In the future research, we will focus on more rigorous multimodal sensor data synchronization methods that minimize the time stamp differences. In addition, we will further study transfer learning method that enables transfer of trained models tailored to the training data to the evaluation data that follows a different distribution. It is expected that a model capable of exhibiting robust recognition performance against changes in data that is not considered in the model learning stage will be obtained.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.45 no.6
• /
• pp.154-161
• /
• 2008

In this paper, a novel method is proposed using color and pattern information for recognizing some emotions included in a fertile. Here we use 10 Kobayashi emotion to represent emotions. - { romantic, clear, natural, casual, elegant chic, dynamic, classic, dandy, modem } The proposed system is composed of feature extraction and classification. To transform the subjective emotions as physical visual features, we extract representative colors and Patterns from textile. Here, the representative color prototypes are extracted by color quantization method, and patterns exacted by wavelet transform followed by statistical analysis. These exacted features are given as input to the neural network (NN)-based classifiers, which decides whether or not a textile had the corresponding emotion. When assessing the effectiveness of the proposed system with 389 textiles collected from various application domains such as interior, fashion, and artificial ones. The results showed that the proposed method has the precision of 100% and the recall of 99%, thereby it can be used in various textile industries.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.05a
• /
• pp.204-206
• /
• 2021

Real-time drilling parameters prediction is a considerably important study from the viewpoint of maximizing drilling efficiency. Among the methods of maximizing drilling, the method of improving the drilling speed is common, which is related to the rate of penetration, drillstring rotational speed, weight on bit, and drilling mud flow rate. This study proposes a method of predicting the drilling rate, one of the real-time drilling parameters, using a recurrent neural network-based deep learning model, and compares the existing physical-based drilling rate prediction model with a prediction model using deep learning.

• Journal of IKEEE
• /
• v.22 no.4
• /
• pp.1123-1130
• /
• 2018

A cat robot for the Autism Spectrum Disorders(ASD) treatment was designed and conducted field test. The designed robot had emotion expressing action through interaction by the touch, and performed a reasonable emotional expression based on Artificial Neural Network(ANN). However these operations were difficult to use in the various healing activities. In this paper, we describe a motion design that can be used in a variety of contexts and flexibly reaction with various kinds of situations. As a necessary element, the speech recognition system using the speech data collection method and ANN was suggested and the classification results were analyzed after experiment. This ANN will be improved through collecting various voice data to raise the accuracy in the future and checked the effectiveness through field test.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.4
• /
• pp.47-53
• /
• 2021

In this paper, we propose a quadtree-based optimization technique that enables fast Super-resolution(SR) computation by efficiently classifying and dividing physics-based simulation data required to calculate SR. The proposed method reduces the time required for quadtree computation by downscaling the smoke simulation data used as input data. By binarizing the density of the smoke in this process, a quadtree is constructed while mitigating the problem of numerical loss of density in the downscaling process. The data used for training is the COCO 2017 Dataset, and the artificial neural network uses a VGG19-based network. In order to prevent data loss when passing through the convolutional layer, similar to the residual method, the output value of the previous layer is added and learned. In the case of smoke, the proposed method achieved a speed improvement of about 15 to 18 times compared to the previous approach.

• Journal of Korea Water Resources Association
• /
• v.51 no.6
• /
• pp.503-514
• /
• 2018

In recent, the hydrological regime of the Mekong river is changing drastically due to climate change and haphazard watershed development including dam construction. Information of hydrologic feature like streamflow of the Mekong river are required for water disaster prevention and sustainable water resources development in the river sharing countries. In this study, runoff simulations at the Kratie station of the lower Mekong river are performed using SWAT (Soil and Water Assessment Tool), a physics-based hydrologic model, and LSTM (Long Short-Term Memory), a data-driven deep learning algorithm. The SWAT model was set up based on globally-available database (topography: HydroSHED, landuse: GLCF-MODIS, soil: FAO-Soil map, rainfall: APHRODITE, etc) and then simulated daily discharge from 2003 to 2007. The LSTM was built using deep learning open-source library TensorFlow and the deep-layer neural networks of the LSTM were trained based merely on daily water level data of 10 upper stations of the Kratie during two periods: 2000~2002 and 2008~2014. Then, LSTM simulated daily discharge for 2003~2007 as in SWAT model. The simulation results show that Nash-Sutcliffe Efficiency (NSE) of each model were calculated at 0.9(SWAT) and 0.99(LSTM), respectively. In order to simply simulate hydrological time series of ungauged large watersheds, data-driven model like the LSTM method is more applicable than the physics-based hydrological model having complexity due to various database pressure because it is able to memorize the preceding time series sequences and reflect them to prediction.

• Journal of the Korea Society of Computer and Information
• /
• v.25 no.5
• /
• pp.19-26
• /
• 2020

This paper proposes a framework for creating acrobatic scenes such as soccer ball lifting using various users' videos. The proposed method can generate a desired result within a few seconds using a general video of user recorded with a mobile phone. The framework of this paper is largely divided into three parts. The first is to analyze the posture by receiving the user's video. To do this, the user can calculate the pose of the user by analyzing the video using a deep learning technique, and track the movement of a selected body part. The second is to analyze the movement trajectory of the selected body part and calculate the location and time of hitting the object. Finally, the trajectory of the object is generated using the analyzed hitting information. Then, a natural object lifting scenes synchronized with the input user's video can be generated. Physical-based optimization was used to generate a realistic moving object. Using the method of this paper, we can produce various augmented reality applications.

• Proceedings of the Korean Society of Disaster Information Conference
• /
• 2017.11a
• /
• pp.339-340
• /
• 2017

화학공장의 누출사고는 초기에 적절히 대응하지 못할 경우 화재 폭발과 같은 2차 3차의 복합재난사고로 확산될 위험성이 매우 높다. 이러한 이유로 누출사고 발생 초기에 누출이 발생한 지점을 신속히 파악하여 현장안전요원에게 알림으로써, 보다 체계적이고 효율적인 초기대응을 가능하게 하여, 사고피해를 완화시킬 수 있는 통합적인 누출사고 대응시스템 구축은 매우 중요하다고 할 수 있다. 본 연구에서는, 통합적인 누출사고 대응시스템 구축을 위한 선행연구로, 딥러닝 기반의 누출원추적 모델 개발을 제안한다. 여수에 위치한 실제 화학공장을 대상으로 누출사고 시나리오에 대한 Computational Fluid Dynamics (CFD) 시뮬레이션을 진행한 뒤, 화학공장 경계면에 배치된 각 센서별 위치에서의 농도, 풍향 그리고 풍속데이터를 추출하고, 센서 좌표를 추가하여 인공신경망을 학습시켰다. 학습된 모델은 40개의 누출후보군에 대해 학습에 사용되지 않은 상황들에서도 75.43%의 정확도로 누출이 일어난 지점을 실시간 예측해냄을 확인하였다. 또한 누출지점 예측이 일치하지 않은 경우도, 예측된 지점이 실제 누출이 일어난 지점과 물리적으로 매우 인접함을 확인함으로써 제안된 모델을 실제 현장에 적용할시 기대되는 효과는 더 클 것으로 판단하였다.