• ETRI Journal
• /
• v.39 no.3
• /
• pp.398-405
• /
• 2017

In this paper, an artificial stereo extension method that creates stereophonic sound from a mono sound source is proposed. The proposed method first trains deep neural networks (DNNs) that model the nonlinear relationship between the dominant and residual signals of the stereo channel. In the training stage, the band-wise log spectral magnitude and unwrapped phase of both the dominant and residual signals are utilized to model the nonlinearities of each sub-band through deep architecture. From that point, stereo extension is conducted by estimating the residual signal that corresponds to the input mono channel signal with the trained DNN model in a sub-band domain. The performance of the proposed method was evaluated using a log spectral distortion (LSD) measure and multiple stimuli with a hidden reference and anchor (MUSHRA) test. The results showed that the proposed method provided a lower LSD and higher MUSHRA score than conventional methods that use hidden Markov models and DNN with full-band processing.

• Journal of Semiconductor Engineering
• /
• v.2 no.2
• /
• pp.130-135
• /
• 2021

Recently, deep neural networks (DNNs) are actively used for action control so that an autonomous system, such as the robot, can perform human-like behaviors and operations. Unlike recognition tasks, the real-time operation is essential in action control, and it is too slow to use remote learning on a server communicating through a network. New learning techniques, such as reinforcement learning (RL), are needed to determine and select the correct robot behavior locally. In this paper, we propose an energy-efficient DNN processor with a LUT-based processing engine and near-zero skipper. A CNN-based facial emotion recognition and an RNN-based emotional dialogue generation model is integrated for natural HRI system and tested with the proposed processor. It supports 1b to 16b variable weight bit precision with and 57.6% and 28.5% lower energy consumption than conventional MAC arithmetic units for 1b and 16b weight precision. Also, the near-zero skipper reduces 36% of MAC operation and consumes 28% lower energy consumption for facial emotion recognition tasks. Implemented in 65nm CMOS process, the proposed processor occupies 1784×1784 um2 areas and dissipates 0.28 mW and 34.4 mW at 1fps and 30fps facial emotion recognition tasks.

• Journal of Multimedia Information System
• /
• v.9 no.3
• /
• pp.177-182
• /
• 2022

A deep neural network (DNN) includes variables whose values keep on changing with the training process until it reaches the final point of convergence. These variables are the co-efficient of a polynomial expression to relate to the feature extraction process. In general, DNNs work in multiple 'dimensions' depending upon the number of channels and batches accounted for training. However, after the execution of feature extraction and before entering the SoftMax or other classifier, there is a conversion of features from multiple N-dimensions to a single vector form, where 'N' represents the number of activation channels. This usually happens in a Fully connected layer (FCL) or a dense layer. This reduced 2D feature is the subject of study for our analysis. For this, we have used the FCL, so the trained weights of this FCL will be used for the weight-class correlation analysis. The popular DNN models selected for our study are ResNet-101, VGG-19, and GoogleNet. These models' weights are directly used for fine-tuning (with all trained weights initially transferred) and scratch trained (with no weights transferred). Then the comparison is done by plotting the graph of feature distribution and the final FCL weights.

• Journal of Broadcast Engineering
• /
• v.28 no.1
• /
• pp.141-144
• /
• 2023

Recently, a joint channel estimation, channel quantization, feedback, and precoding system based on deep-neural network (DNN) was proposed. The corresponding system achieved a joint optimization based on deep learning such that it achieved a higher sum rate than the existing codebook-based precoding systems. However, this DNN-based procoding system is not directly applicable for the environments with many users such that a specific user selection can potentially increase the sum rate of the system. Thus, in this letter, we study an appropriate user selection method suitable for DNN-based precoding.

• Journal of Korea Multimedia Society
• /
• v.23 no.3
• /
• pp.402-411
• /
• 2020

Power line noise in electrocardiogram signals makes it difficult to diagnose cardiovascular disease. ECG signals without power line noise are needed to increase the accuracy of diagnosis. In this paper, it is proposed DNN(Deep Neural Network) model to remove the power line noise in ECG. The proposed model is learned with noisy ECG, and clean ECG. Performance of the proposed model were performed in various environments(varying amplitude, frequency change, real-time amplitude change). The evaluation used signal-to-noise ratio and root mean square error (RMSE). The difference in evaluation metrics between the noisy ECG signals and the de-noising ECG signals can demonstrate effectiveness as the de-noising model. The proposed DNN model learning result was a decrease in RMSE 0.0224dB and a increase in signal-to-noise ratio 1.048dB. The results performed in various environments showed a decrease in RMSE 1.7672dB and a increase in signal-to-noise ratio 15.1879dB in amplitude changes, a decrease in RMSE 0.0823dB and a increase in signal-to-noise ratio 4.9287dB in frequency changes. Finally, in real-time amplitude changes, RMSE was decreased 0.3886dB and signal-to-noise ratio was increased 11.4536dB. Thus, it was shown that the proposed DNN model can de-noise power line noise in ECG.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.41 no.8
• /
• pp.958-964
• /
• 2016

In this paper, we propose an i-vector based language recognition system to identify the spoken language of the speaker, which uses multiple discriminative deep neural network (DNN) models analogous to the multi-class support vector machine (SVM) classification system. The proposed model was trained and tested using the i-vectors included in the NIST 2015 i-vector Machine Learning Challenge database, and shown to outperform the conventional language recognition methods such as cosine distance, SVM and softmax NN classifier in open-set experiments.

• The Journal of Engineering Geology
• /
• v.31 no.1
• /
• pp.31-42
• /
• 2021

Database of Cut-slope management system (CSMS) has been constructed based on investigations of all slopes on the roads of the whole country. The investigation data is documented by human, so it is inevitable to avoid human-error such as missing-data and incorrect entering data into computer. The goal of this paper is constructing a prediction model based on several machine-learning algorithms to solve those imperfection problems of the CSMS data. First of all, the character-type data in CSMS data must be transformed to numeric data. After then, two algorithms, i.g., multinomial logistic regression and deep-neural-network (DNN), are performed, and those prediction models from two algorithms are compared. Finally, it is identified that the accuracy of DNN-model is better than logistic model, and the DNN-model will be utilized to improve data-quality.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.16 no.3
• /
• pp.123-129
• /
• 2023

if we know what we're looking at, we can get a lot of information. Due to the development of eye tracking, Information on gaze point can be obtained through software provided by various eye tracking equipments. However, it is difficult to estimate accurate information such as the actual gaze depth. If it is possible to calibrate the eye tracker with the actual gaze depth, it will enable the derivation of realistic and accurate results with reliable validity in various fields such as simulation, digital twin, VR, and more. Therefore, in this paper, we experiment with acquiring and calibrating raw gaze depth using an eye tracker and software. The experiment involves designing a Deep Neural Network (DNN) model and then acquiring gaze depth values provided by the software for specified distances from 300mm to 10,000mm. The acquired data is trained through the designed DNN model and calibrated to correspond to the actual gaze depth. In our experiments with the calibrated model, we were able to achieve actual gaze depth values of 297mm, 904mm, 1,485mm, 2,005mm, 3,011mm, 4,021mm, 4,972mm, 6,027mm, 7,026mm, 8,043mm, 9,021mm, and 10,076mm for the specified distances from 300mm to 10,000mm.

• Journal of Biomedical Engineering Research
• /
• v.45 no.3
• /
• pp.128-137
• /
• 2024

It is necessary to develop a pulsatile Extracorporeal Membrane Oxygenator (p-ECMO) with counter-pulsation control(CPC), which ejects blood during the diastolic phase of the heart rather than the systolic phase, due to the known issues with conventional ECMO causing fatal complications such as ventricular dilation and pulmonary edema. A promising method to simultaneously detect the pulsations of the heart and p-ECMO is to analyze blood pressure waveforms using deep neural network technology(DNN). However, the accurate detection of cardiac rhythms by DNNs is challenging due to various noises such as pulsations from p-ECMO, reflected waves in the vessels, and other dynamic noises. This study aims to evaluate the accuracy of DNNs developed for CPC in p-ECMO, using human-like blood pressure waveforms reproduced in an in-vitro experiment. Especially, an experimental setup that reproduces reflected waves commonly observed in actual patients was developed, and the impact of these waves on DNN judgments was assessed using a multiple DNN (m-DNN) that provides accurate determinations along with a separate index for heartbeat recognition ability. In the experimental setup inducing reflected waves, it was observed that the shape of the blood pressure waveform became increasingly complex, which coincided with an increase in harmonic components, as evident from the Fast Fourier Transform results of the blood pressure wave. It was observed that the recognition score (RS) of DNNs decreased in blood pressure waveforms with significant harmonic components, separate from the frequency components caused by the heart and p-ECMO. This study demonstrated that each DNN trained on blood pressure waveforms without reflected waves showed low RS when faced with waveforms containing reflected waves. However, the accuracy of the final results from the m-DNN remained high even in the presence of reflected waves.

• International Journal of Internet, Broadcasting and Communication
• /
• v.12 no.4
• /
• pp.26-32
• /
• 2020

There are many airline customer evaluation data, but they are insufficient in terms of predicting customer satisfaction in practice. In particular, they are generally insufficient in case of verification of data value and development of a customer satisfaction prediction model based on customer evaluation data. In this paper, airline customer satisfaction analysis is conducted through an experiment of correlation analysis between customer evaluation data provided by Google's Kaggle. The difference in accuracy varied according to the three types, which are the overall variables, the top 4 and top 8 variables with the highest correlation. To build an airline customer satisfaction prediction model, they are applied to three classification algorithms of Random Forest, SVM, DNN and conduct a classification experiment. They are divided into training data and verification data by 7:3. As a result, the DNN model showed the lowest accuracy at 86.4%, while the SVM model at 89% and the Random Forest model at 95.7% showed the highest accuracy and performance.