• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.28 no.12C
• /
• pp.1250-1257
• /
• 2003

Generally images are classified into two regions: edge and flat region. While low frequency components are popular in the flat region, high frequency components are quite important in the edge region. Therefore, deinterlacing algorithm that considers the characteristic of each region can be more efficient. In this paper, an image is divided into edge region and flat region by the local variance. And then, for each region, frequency modularized neural network is assigned. Using this structure, each modularized neural network can learn only its region intensively and avoid the complexity of learning caused by the data of different region. Using the local AC data for the input of neural network can prevent the degradation of the performance of teaming due to the average intensity values of image that disturbs the effective learning. The proposed method shows the improved performance compared with previous algorithms in the simulation.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.32B no.1
• /
• pp.171-179
• /
• 1995

In many continuous speech recognition systems, phoneme is used as a basic recognition unit However, the coarticulation generated among neighboring phonemes makes difficult to recognize phonemes consistently. This paper proposes allophone as an alternative recognition unit. We have classified each phoneme into three different allophone groups by the location of phoneme within a syllable. For a recognition algorithm, time-delay neural network(TDNN) has been designed. To recognize all Korean allophones, TDNNs are constructed in modular fashion according to acoustic-phonetic features (e.g. voiced/unvoiced, the location of phoneme within a word). Each TDNN is trained independently, and then they are integrated hierarchically into a whole speech recognition system. In this study, we have experimented Korean plosives with phoneme-based recognition system and allophone-based recognition system. Experimental results show that allophone-based recognition is much less affected by the coarticulation.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2001.10a
• /
• pp.555-558
• /
• 2001

본 논문에서는 인간 행동의 성별 인식문제를 해결하기 위해 여러 개의 전문가(expert) 신경망의 앙상블로 이루어진 결합 신경망 분류기를 제안한다. 하나는 여러 개의 modular 다층퍼셉트론을 계층형으로 결합한 모텔이고, 다른 하나는 modular 다층퍼셉트론들의 출력값을 의사결정트리로 결합하는 모델이다. 데이터 베이스는 남녀 각 13 명의 데이터로 이루어져 있고, 문 두드리기, 손 흔들기, 물건 들어올리기의 세 가지 동작을, 보통 상태 혹은 화난 상태하에서 10 회씩 반복 수행하여 저장하였다. 행위자의 움직임은 몸에 부착된 6 개의 적외선 센서를 사용하여 기록 되었으며, 2 차원 혹은 3 차원 속도 및 좌표가 그 특징값으로 사용되었다. 앙상블 분류기의 성능을 비교하기 위하여 단일 다층퍼셉트론, 의사결정트리, 자기구성지도 및 support vector machine 을 사용한 실험 결과를 보였다. 실험 결과, 신경망 앙상블 모델이 다른 전통적인 분류기 및 사람에 비하여 훨씬 우수한 성능을 보였음을 알 수 있었다.

• Proceedings of the IEEK Conference
• /
• 2002.06c
• /
• pp.127-130
• /
• 2002

Human's brain action can divide by recognition and intelligence. recognition is sensing voice, image and smell and Intelligence is logical judgment, inference, decision. To this concept, Define function of cerebral cortex, and apply the result. Current expert system is lack, that reasoning by cerebral cortex and thalamus, hoppocampal and so on. In this paper, With human's brain action, wish to embody human's action artificially Embody brain mechanism using Modular Neural Network, Applied this result to snake robot.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1991.06a
• /
• pp.11-14
• /
• 1991

음소단위로 구성된 음소군들 각각에 대해 구성된 신경 회로망을 하나로 통합하는 모듈구조로 신경망을 이용하여 일반적인 예약 시스템에서 사용할 수 있는 어휘인 시간명, 월명, 지역명등 총 34 단어에 대한 인식 실험내용을 기술한다. 구문회로망(context net)를 이용하는 경우에 약 91.2%의 인식율을, 단순히 음소단위를 기반으로하여 인식할 경우에 약 72%의 인식율을 얻으므로써, 음소 단위 인식시스템의 경우에 보다 높은 인식율을 얻기 위해서는 상위 level의 처리가 수반되어야 함을 확인할 수 있었다.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2002.05a
• /
• pp.256-259
• /
• 2002

Recently, most vehicles has the Automatic transmission system as their transmission system. The automatic transmission system operates with fixed shift patterns. In the opposite of manual operation, it is easy and convenient for driving. Though these merit, the system can not evaluate the driver's intension because of usage of firmed shift pattern. Especially, when the load has declination the AT system must operate for engine break effect. Namely, if the vehicle drives on the load of decrease, the acceleration of the vehicle goes to high then. At that time, the shift goes to down position the vehicle has some negative acceleration with the resistance of engine. To consider driver's intension in this case, we must consider both the driving intensity of driver and the status of load. In this paper, we developed flexible automatic transmission system by using the proposed moduled neural networks which can learn the status of the load and driver's intensity As a result, we compare the transmission system using firmed shift pattern and the proposed transmission system and show the good performance in the change of shift position.

• The Transactions of the Korea Information Processing Society
• /
• v.4 no.7
• /
• pp.1881-1892
• /
• 1997

This paper proposes the method for recognition of printed chinese characters by probabilistic VQ networks and multi-stage recognizer has hierarchical structure. We use modular neural networks, because it is difficult to construct a large-scale neural network. Problems in this procedure are replaced by probabilistic neural network model. And, Confused Characters which have significant ratio of miss-classification are reclassified using the entropy theory. The experimental object consists of 4,619 chinese characters within the KSC5601 code except the same shape but different code. We have 99.33% recognition rate to the training data, and 92.83% to the test data. And, the recognition speed of system is 4-5 characters per second. Then, these results demonstrate the usefulness of our work.

• Environmental Engineering Research
• /
• v.15 no.2
• /
• pp.123-126
• /
• 2010

Modeling non-point pollution across multiple scales has become an important environmental issue. As a more representative and practical approach in quantifying and qualifying surface water, a modular neural network (MNN) was implemented in this study. Two different site-scales ($1.5\;{\times}\;10^5$ and $1.62\;{\times}\;10^6\;m^2$) with the same plants, soils, and paddy field management practices, were selected. Hydrologic data (rainfall, irrigation and surface discharge) and water quality data (time-series nutrient loadings) were continuously monitored and then used for the verification of MNN performance. Correlation coefficients (R) for the results predicted from the networks versus measured values were within the range of 0.41 to 0.95. The small block could be extrapolated to the large field for the rainfall-surface drainage process. Nutrient prediction produced less favorable results due to the complex phenomena of nutrients in the drainage water. However, the feasibility of using MNN to generate improved prediction accuracy was demonstrated if more hydrologic and environmental data are provided. The study findings confirmed the estimation accuracy of the upscaling from a small-segment block to large-scale paddy field, thereby contributing to the establishment of water quality management for sustainable agriculture.

• Proceedings of the KIEE Conference
• /
• 2007.07a
• /
• pp.1713-1715
• /
• 2007

Measurement of network traffic have shown that the self-similarity is a ubiquitous phenomenon spanning across diverse network environments. In previous work, we have explored the feasibility of exploiting the long-range correlation structure in a self-similar traffic for the congestion control. We have advanced the framework of the multiple time scale congestion control and showed its effectiveness at enhancing performance for the rate-based feedback control. Our contribution is threefold. First, we define a modular extension of the TCP-a function called with a simple interface-that applies to various flavours of the TCP-e.g., Tahoe, Reno, Vegas and show that it significantly improves performance. Second, we show that a multiple time scale TCP endows the underlying feedback control with proactivity by bridging the uncertainty gap associated with reactive controls which is exacerbated by the high delay-bandwidth product in broadband wide area networks. Third, we investigate the influence of the three traffic control dimensions-tracking ability, connection duration, and fairness-on performance.

• Nuclear Engineering and Technology
• /
• v.51 no.3
• /
• pp.723-730
• /
• 2019

Acquiring instrumentation signals generated from nuclear power plants (NPPs) is essential to maintain nuclear reactor integrity or to mitigate an abnormal state under normal operating conditions or severe accident circumstances. However, various safety-critical instrumentation signals from NPPs cannot be accurately measured on account of instrument degradation or failure under severe accident circumstances. Reactor vessel (RV) water level, which is an accident monitoring variable directly related to reactor cooling and prevention of core exposure, was predicted by applying a few signals to deep neural networks (DNNs) during severe accidents in NPPs. Signal data were obtained by simulating the postulated loss-of-coolant accidents at hot- and cold-legs, and steam generator tube rupture using modular accident analysis program code as actual NPP accidents rarely happen. To optimize the DNN model for RV water level prediction, a genetic algorithm was used to select the numbers of hidden layers and nodes. The proposed DNN model had a small root mean square error for RV water level prediction, and performed better than the cascaded fuzzy neural network model of the previous study. Consequently, the DNN model is considered to perform well enough to provide supporting information on the RV water level to operators.