• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.14 no.11
• /
• pp.4290-4309
• /
• 2020

In recent years, advancements in machine learning capabilities have allowed it to see widespread adoption for tasks such as object detection, image classification, and anomaly detection. However, despite their promise, a limitation lies in the fact that a network's performance quality is based on the data which it receives. A well-trained network will still have poor performance if the subsequent data supplied to it contains artifacts, out of focus regions, or other visual distortions. Under normal circumstances, images of the same scene captured from differing points of focus, angles, or modalities must be separately analysed by the network, despite possibly containing overlapping information such as in the case of images of the same scene captured from different angles, or irrelevant information such as images captured from infrared sensors which can capture thermal information well but not topographical details. This factor can potentially add significantly to the computational time and resources required to utilize the network without providing any additional benefit. In this study, we plan to explore using image fusion techniques to assemble multiple images of the same scene into a single image that retains the most salient key features of the individual source images while discarding overlapping or irrelevant data that does not provide any benefit to the network. Utilizing this image fusion step before inputting a dataset into the network, the number of images would be significantly reduced with the potential to improve the classification performance accuracy by enhancing images while discarding irrelevant and overlapping regions.

• Journal of KIISE:Software and Applications
• /
• v.27 no.5
• /
• pp.463-472
• /
• 2000

In this paper, we consider the problem of realizing associative memories via cellular neural network(CNN). After introducing qualitative properties of the CNN model, we formulate the synthesis of CNN that can store given binary vectors with optimal performance as a constrained optimization problem. Next, we observe that this problem's constraints can be transformed into simple inequalities involving linear matrix inequalities(LMIs). Finally, we reformulate the synthesis problem as a generalized eigenvalue problem(GEVP), which can be efficiently solved by recently developed interior point methods. Proposed method can be applied to both space varying template CNNs and space-invariant template CNNs. The validity of the proposed approach is illustrated by design examples.

• Journal of KIISE
• /
• v.43 no.12
• /
• pp.1356-1364
• /
• 2016

This paper proposes a novel aesthetic photo recomposition method using a deep convolutional neural network (DCNN). Previous recomposition approaches define the aesthetic score of photo composition based on the distribution of salient objects, and enhance the photo composition by maximizing the score. These methods suffer from heavy computational overheads, and often fail to enhance the composition because their optimization depends on the performance of existing salient object detection algorithms. Unlike previous approaches, we address the photo recomposition problem by utilizing DCNN, which shows remarkable performance in object detection and recognition. DCNN is used to iteratively predict cropping directions for a given photo, thus generating an aesthetically enhanced photo in terms of composition. Experimental results and user study show that the proposed framework can automatically crop the photo to follow specific composition guidelines, such as the rule of thirds.

• Journal of IKEEE
• /
• v.23 no.4
• /
• pp.1295-1301
• /
• 2019

In this paper, we propose a novel method of denoising human motion using a bidirectional recurrent neural network (BRNN) with an attention mechanism. The corrupted motion captured from a single 3D depth sensor camera is automatically fixed in the well-established smooth motion manifold. Incorporating an attention mechanism into BRNN achieves better optimization results and higher accuracy than other deep learning frameworks because a higher weight value is selectively given to a more important input pose at a specific frame for encoding the input motion. Experimental results show that our approach effectively handles various types of motion and noise, and we believe that our method can sufficiently be used in motion capture applications as a post-processing step after capturing human motion.

• The Journal of the Acoustical Society of Korea
• /
• v.38 no.5
• /
• pp.607-613
• /
• 2019

This paper proposes LF-MMI (Lattice Free Maximum Mutual Information)-based acoustic modeling using external knowledge for speech recognition. Note that an external knowledge refers to text data other than training data used in acoustic model. LF-MMI, objective function for optimization of training DNN (Deep Neural Network), has high performances in discriminative training. In LF-MMI, a phoneme probability as prior probability is used for predicting posterior probability of the DNN-based acoustic model. We propose using external knowledges for training the prior probability model to improve acoustic model based on DNN. It is measured to relative improvement 14 % as compared with the conventional LF-MMI-based model.

• Journal of IKEEE
• /
• v.23 no.1
• /
• pp.134-142
• /
• 2019

As the development of internet of energy (IoE) technologies and spread of various electronic devices have diversified patterns of energy consumption, the reliability of demand prediction has decreased, causing problems in optimization of power generation and stabilization of power supply. In this study, we propose a deep learning method, 1-Dimention-Convolution and Bidirectional Long Short-Term Memory (1D-ConvBLSTM), that combines a convolution neural network (CNN) and a Bidirectional Long Short-Term Memory(BLSTM) for highly reliable demand forecasting by effectively extracting the energy consumption pattern. In experimental results, the demand is predicted with the proposed deep learning method for various number of learning iterations and feature maps, and it is verified that the test data is predicted with a small number of iterations.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.16 no.4
• /
• pp.625-632
• /
• 2021

This paper studies the effect of the number of filters in the CNN (Convolutional Neural Network) layer on the performance of a noise attenuator. Speech is estimated from a noised speech signal using a 64-neuron, 16-kernel CNN filter and an error back-propagation algorithm. In this study, in order to verify the performance of the noise attenuator with respect to the number of filters, a program using Keras library was written and simulation was performed. As a result of simulation, it can be seen that this system has the smallest MSE (Mean Squared Error) and MAE (Mean Absolute Error) values when the number of filters is 16, and the performance is the lowest when there are 4 filters. And when there are more than 8 filters, it was shown that the MSE and MAE values do not differ significantly depending on the number of filters. From these results, it can be seen that about 8 or more filters must be used to express the characteristics of the speech signal.

• Smart Structures and Systems
• /
• v.29 no.1
• /
• pp.117-127
• /
• 2022

Data anomalies seriously threaten the reliability of the bridge structural health monitoring system and may trigger system misjudgment. To overcome the above problem, an efficient and accurate data anomaly detection method is desiderated. Traditional anomaly detection methods extract various abnormal features as the key indicators to identify data anomalies. Then set thresholds artificially for various features to identify specific anomalies, which is the artificial experience method. However, limited by the poor generalization ability among sensors, this method often leads to high labor costs. Another approach to anomaly detection is a data-driven approach based on machine learning methods. Among these, the bidirectional long-short memory neural network (BiLSTM), as an effective classification method, excels at finding complex relationships in multivariate time series data. However, training unprocessed original signals often leads to low computation efficiency and poor convergence, for lacking appropriate feature selection. Therefore, this article combines the advantages of the two methods by proposing a deep learning method with manual experience statistical features fed into it. Experimental comparative studies illustrate that the BiLSTM model with appropriate feature input has an accuracy rate of over 87-94%. Meanwhile, this paper provides basic principles of data cleaning and discusses the typical features of various anomalies. Furthermore, the optimization strategies of the feature space selection based on artificial experience are also highlighted.

• Journal of Korean Society of Industrial and Systems Engineering
• /
• v.45 no.4
• /
• pp.233-239
• /
• 2022

The injection molding process is a process in which thermoplastic resin is heated and made into a fluid state, injected under pressure into the cavity of a mold, and then cooled in the mold to produce a product identical to the shape of the cavity of the mold. It is a process that enables mass production and complex shapes, and various factors such as resin temperature, mold temperature, injection speed, and pressure affect product quality. In the data collected at the manufacturing site, there is a lot of data related to good products, but there is little data related to defective products, resulting in serious data imbalance. In order to efficiently solve this data imbalance, undersampling, oversampling, and composite sampling are usally applied. In this study, oversampling techniques such as random oversampling (ROS), minority class oversampling (SMOTE), ADASYN(Adaptive Synthetic Sampling), etc., which amplify data of the minority class by the majority class, and complex sampling using both undersampling and oversampling, are applied. For composite sampling, SMOTE+ENN and SMOTE+Tomek were used. Artificial neural network techniques is used to predict product quality. Especially, MLP and RNN are applied as artificial neural network techniques, and optimization of various parameters for MLP and RNN is required. In this study, we proposed an SA technique that optimizes the choice of the sampling method, the ratio of minority classes for sampling method, the batch size and the number of hidden layer units for parameters of MLP and RNN. The existing sampling methods and the proposed SA method were compared using accuracy, precision, recall, and F1 Score to prove the superiority of the proposed method.

• Journal of Powder Materials
• /
• v.29 no.6
• /
• pp.459-467
• /
• 2022

Soft magnetic powder materials are used throughout industries such as motors and power converters. When manufacturing Fe-based soft magnetic composites, the size and shape of the soft magnetic powder and the microstructure in the powder are closely related to the magnetic properties. In this study, Fe-Si-Al-P alloy powders were manufactured using various manufacturing process parameter sets, and the process parameters of the vacuum induction melt gas atomization process were set as melt temperature, atomization gas pressure, and gas flow rate. Process variable data that records are converted into 6 types of data for each powder recovery section. Process variable data that recorded minute changes were converted into 6 types of data and used as input variables. As output variables, a total of 6 types were designated by measuring the particle size, flowability, apparent density, and sphericity of the manufactured powders according to the process variable conditions. The sensitivity of the input and output variables was analyzed through the Pearson correlation coefficient, and a total of 6 powder characteristics were analyzed by artificial neural network model. The prediction results were compared with the results through linear regression analysis and response surface methodology, respectively.