• Journal of Broadcast Engineering
• /
• v.27 no.3
• /
• pp.332-340
• /
• 2022

Recently, with the development of deep learning and artificial neural network technologies, research on the application of neural network has been actively conducted in the field of video coding. In particular, deep learning-based intra prediction is being studied as a way to overcome the performance limitations of the existing intra prediction techniques. This paper presents a method of context-adaptive neural network-based intra prediction model training and its coding performance analysis. In other words, in this paper, we implement and train a known intra prediction model based on convolutional neural network (CNN) that predicts a current block using contextual information from reference blocks. Then, we integrate the trained model into HM16.19 as an additional intra prediction mode and evaluate the coding performance of the trained model. Experimental results show that the trained model gives 0.28% BD-rate bit saving over HEVC in All Intra (AI) coding mode. In addition, the coding performance change of training considering block partition is also presented.

• 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.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.32B no.12
• /
• pp.1663-1669
• /
• 1995

This paper is a study on the recognition of Korean proverb using neural network and Markov model. The neural network uses, at the stage of training neurons, features such as the rate of zero crossing, short-term energy and PLP-Cepstrum, covering a time of 300ms long. Markov models were generated by the recognized phoneme strings. The recognition of words and proverbs using Markov models have been carried out. Experimental results show that phoneme and word recognition rates are 81. 2%, 94.0% respectively for Korean proverb recognition experiments.

• Journal of Korea Water Resources Association
• /
• v.34 no.6
• /
• pp.701-711
• /
• 2001

Forecasting of water quality variation is not an easy process due to the complicated nature of various water quality factors and their interrelationships. The objective of this study is to test the applicability of neural network models to the forecasting of the water quality at Gongju station in Geum River. This is done by forecasting monthly water qualities such as DO, BOD, and TN, and comparing with those obtained by ARIMA model. The neural network models of this study use BP(Back Propagation) algorithm for training. In order to improve the performance of the training, the models are tested in three different styles ; MANN model which uses the Moment-Adaptive learning rate method, LMNN model which uses the Levenberg-Marquardt method, and MNN model which separates the hidden layers for judgement factors from the hidden layers for water quality data. the results show that the forecasted water qualities are reasonably close to the observed data. And the MNN model shows the best results among the three models tested

• Journal of the Korean Society for Nondestructive Testing
• /
• v.24 no.4
• /
• pp.362-370
• /
• 2004

In this work, the collapse moment due to wall-thinning defects is estimated by using fuzzy neural networks. The developed fuzzy neural networks have been applied to the numerical data obtained from the finite element analysis. Principal component analysis is used to preprocess the input signals into the fuzzy neural network to reduce the sensitivity to the input change and the fuzzy neural networks are trained by using the data set prepared for training (training data) and verified by using another data set different (independent) from the training data. Also, two fuzzy neural networks are trained for two data sets divided into the two classes of extrados and intrados defects, which is because they have different characteristics. The relative 2-sigma errors of the estimated collapse moment are 3.07% for the training data and 4.12% for the test data. It is known from this result that the fuzzy neural networks are sufficiently accurate to be used in the wall-thinning monitoring of elbows.

• Journal of the Korean Institute of Telematics and Electronics S
• /
• v.36S no.1
• /
• pp.29-37
• /
• 1999

In general, the information processing capability of a neural network is determined by its architecture and efficient training patterns. However, there is no systematic method for designing neural network and selecting effective training patterns. Evolutionary Algorithms(EAs) are referred to as the methods of population-based optimization. Therefore, EAs are considered as very efficient methods of optimal system design because they can provide much opportunity for obtaining the global optimal solution. In this paper, we propose a new method for finding the optimal structure of neural networks based on competitive co-evolution, which has two different populations. Each population is called the primary population and the secondary population respectively. The former is composed of the architecture of neural network and the latter is composed of training patterns. These two populations co-evolve competitively each other, that is, the training patterns will evolve to become more difficult for learning of neural networks and the architecture of neural networks will evolve to learn this patterns. This method prevents the system from the limitation of the performance by random design of neural networks and inadequate selection of training patterns. In co-evolutionary method, it is difficult to monitor the progress of co-evolution because the fitness of individuals varies dynamically. So, we also introduce the measurement method. The validity and effectiveness of the proposed method are inspected by applying it to the visual servoing of robot manipulators.

• Journal of Korea Multimedia Society
• /
• v.23 no.8
• /
• pp.977-985
• /
• 2020

Aiming at the problem of complex feature extraction and low accuracy in human action recognition, this paper proposed a network structure combining batch normalization algorithm with GoogLeNet network model. Applying Batch Normalization idea in the field of image classification to action recognition field, it improved the algorithm by normalizing the network input training sample by mini-batch. For convolutional network, RGB image was the spatial input, and stacked optical flows was the temporal input. Then, it fused the spatio-temporal networks to get the final action recognition result. It trained and evaluated the architecture on the standard video actions benchmarks of UCF101 and HMDB51, which achieved the accuracy of 93.42% and 67.82%. The results show that the improved convolutional neural network has a significant improvement in improving the recognition rate and has obvious advantages in action recognition.

• ETRI Journal
• /
• v.39 no.6
• /
• pp.841-850
• /
• 2017

We investigate neural network image reconstruction for magnetic particle imaging. The network performance strongly depends on the convolution effects of the spectrum input data. The larger convolution effect appearing at a relatively smaller nanoparticle size obstructs the network training. The trained single-layer network reveals the weighting matrix consisting of a basis vector in the form of Chebyshev polynomials of the second kind. The weighting matrix corresponds to an inverse system matrix, where an incoherency of basis vectors due to low convolution effects, as well as a nonlinear activation function, plays a key role in retrieving the matrix elements. Test images are well reconstructed through trained networks having an inverse kernel matrix. We also confirm that a multi-layer network with one hidden layer improves the performance. Based on the results, a neural network architecture overcoming the low incoherence of the inverse kernel through the classification property is expected to become a better tool for image reconstruction.

• Proceedings of the Korean Society for Noise and Vibration Engineering Conference
• /
• 2007.05a
• /
• pp.978-982
• /
• 2007

A new neuro-control scheme for active control of structures is proposed. It utilizes lattice pattern of state vector as training data of probabilistic neural network (PNN). Therefore, it is the so-called lattice probabilistic neural network (LPNN). PNN makes control forces by using all the training patterns. Therefore, it takes much time to obtain a control force in application. This inevitably may delay the control action. However, control force of LPNN is calculated by using only the adjacent information of LPNN input. So, the response of LPNN is greatly faster than PNN. The proposed control algorithm is applied for one story building under California and El Centro earthquakes. Also, control results of the LPNN are compared with those of the conventional PNN. The structural responses have been suppressed effectively by the proposed algorithm.

• The Korean Journal of Ecology
• /
• v.23 no.2
• /
• pp.89-100
• /
• 2000

On community data. sampled in regular intervals on a long-term basis. artificial neural networks were implemented to extract information on characterizing patterns of community changes. The Adaptive Resonance Theory and Kohonen Network were both utilized in learning benthic macroinvertebrate communities in the Soktae Stream of the Suyong River collected monthly for three years. Initially, by regarding each monthly collection as a separate sample unit, communities were grouped into similar patterns after training with the networks. Subsequently, changes in communities in a sequence of samplings (e.g., two-month, four-month, etc.) were given as input to the networks. After training, it was possible to recognize new data set in line with the sampling procedure. Through the comparative study on benthic macroinvertebrates with these learning processes, patterns of community changes in chironomids diverged while those of the total benthic macro-invertebrates tended to be more stable.