• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.11a
• /
• pp.102-105
• /
• 2019

본 논문에서는 심층 신경망 검색 방법을 사용하여 이미지 고해상도화를 위한 심층 신경망을 설계하는 방법을 구현하였다. 일반적으로 이미지 고해상도화, 잡음 제거 및 번짐 제거를 위한 심층신경망 구조는 사람이 설계하였다. 최근에는 이미지 분류 등 다른 영상처리 기법에서 사용하는 심층 신경망 구조를 검색하기 위한 방법이 연구되었다. 본 논문에서는 강화학습을 사용하여 이미지 고해상도화를 위한 심층 신경망 구조를 검색하는 방법을 제안하였다. 제안된 방법은 policy gradient 방법의 일종인 REINFORCE 알고리즘을 사용하여 심층 신경망 구조를 출력하여 주는 제어용 RNN(recurrent neural network)을 학습하고, 최종적으로 이미지 고해상도화를 잘 실현할 수 있는 심층 신경망 구조를 검색하여 설계하였다. 제안된 심층 신경망 구조를 사용하여 이미지 고해상도화를 구현하였고, 약 36.54dB 의 피크 신호 대비 잡음 비율(PSNR)을 가지는 것을 확인할 수 있었다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.125-130
• /
• 2023

최근 심층 신경망을 활용한 한국어 자연어 처리에 대한 관심이 높아지고 있지만, 한국어 자연어 처리에 적합한 신경망 구조 탐색에 대한 연구는 이뤄지지 않았다. 본 논문에서는 문서 분류 정확도를 보상으로 하는 강화 학습 알고리즘을 이용하여 장단기 기억 신경망으로 한국어 문서 분류에 적합한 심층 신경망 구조를 탐색하였으며, 탐색을 위해 사전 학습한 한국어 임베딩 성능과 탐색한 신경망 구조를 분석하였다. 탐색을 통해 찾아낸 신경망 구조는 기존 한국어 자연어 처리 모델에 대해 4 가지 한국어 문서 분류 과제로 비교하였을 때 일반적으로 성능이 우수하고 모델의 크기가 작아 효율적이었다.

• Korean Journal of Remote Sensing
• /
• v.39 no.5_1
• /
• pp.695-713
• /
• 2023

In coastal waters, phytoplankton,suspended particulate matter, and dissolved organic matter intricately and nonlinearly alter the reflectivity of seawater. Neural network technology, which has been rapidly advancing recently, offers the advantage of effectively representing complex nonlinear relationships. In previous studies, a three-stage neural network was constructed to extract the inherent optical properties of each component. However, this study proposes an algorithm that directly employs a deep neural network. The dataset used in this study consists of synthetic data provided by the International Ocean Color Coordination Group, with the input data comprising above-surface remote-sensing reflectance at nine different wavelengths. We derived inherent optical properties using this dataset based on a deep neural network. To evaluate performance, we compared it with a quasi-analytical algorithm and analyzed the impact of log transformation on the performance of the deep neural network algorithm in relation to data distribution. As a result, we found that the deep neural network algorithm accurately estimated the inherent optical properties except for the absorption coefficient of suspended particulate matter (R² greater than or equal to 0.9) and successfully separated the sum of the absorption coefficient of suspended particulate matter and dissolved organic matter into the absorption coefficient of suspended particulate matter and dissolved organic matter, respectively. We also observed that the algorithm, when directly applied without log transformation of the data, showed little difference in performance. To effectively apply the findings of this study to ocean color data processing, further research is needed to perform learning using field data and additional datasets from various marine regions, compare and analyze empirical and semi-analytical methods, and appropriately assess the strengths and weaknesses of each algorithm.

• The Journal of the Acoustical Society of Korea
• /
• v.38 no.3
• /
• pp.308-313
• /
• 2019

Parkinson's disease, one of the three major diseases in old age, has more than 70 % of patients with speech disorders, and recently, diagnostic methods of Parkinson's disease through speech signals have been devised. In this paper, we propose a method of diagnosis of Parkinson's disease based on deep residual gated recurrent neural network using speech features. In the proposed method, the speech features for diagnosing Parkinson's disease are selected and applied to the deep residual gated recurrent neural network to classify Parkinson's disease patients. The proposed deep residual gated recurrent neural network, an algorithm combining residual learning with deep gated recurrent neural network, has a higher recognition rate than the traditional method in Parkinson's disease diagnosis.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.18 no.3
• /
• pp.106-118
• /
• 2019

In recent years, research and development on image object recognition using artificial intelligence have been actively carried out, and it is expected to be used for road maintenance. Among them, artificial intelligence models for object detection of road surface are continuously introduced. In order to develop such object recognition algorithms, a backbone network that extracts feature maps is essential. In this paper, we will discuss how to select the appropriate neural network. To accomplish it, we compared with 4 different deep neural networks using 6,000 road surface damage images. Based on three evaluation methods for analyzing characteristics of neural networks, we propose a method to determine optimal neural networks. In addition, we improved the performance through optimal tuning of hyper-parameters, and finally developed a light backbone network that can achieve 85.9% accuracy of road surface damage classification.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.21 no.12
• /
• pp.2291-2297
• /
• 2017

This paper describe to extract speech measure algorithm for evaluating a speech database, and presents generating method of a speech quality measure using DNN(Deep Neural Network). In our previous study, to produce an effective speech quality measure, we propose a combination of various speech measures which are highly correlated with WER(Word Error Rate). The new combination of various types of speech quality measures in this study is more effective to predict the speech recognition performance compared to each speech measure alone. In this paper, we describe the method of extracting measure using DNN, and we change one of the combined measure from GMM(Gaussican Mixture Model) score used in the previous study to DNN score. The combination with DNN score shows a higher correlation with WER compared to the combination with GMM score.

• Smart Media Journal
• /
• v.5 no.3
• /
• pp.35-41
• /
• 2016

In this paper, we propose hand-gesture signal recognition based on EPS(Electronic Potential Sensor) using Deep learning model. Extracted signals which from Electronic field based sensor, EPS have much of the noise, so it must remove in pre-processing. After the noise are removed with filter using frequency feature, the signals are reconstructed with dimensional transformation to overcome limit which have just one-dimension feature with voltage value for using convolution operation. Then, the reconstructed signal data is finally classified and recognized using multiple learning layers model based on deep learning. Since the statistical model based on probability is sensitive to initial parameters, the result can change after training in modeling phase. Deep learning model can overcome this problem because of several layers in training phase. In experiment, we used two different deep learning structures, Convolutional neural networks and Recurrent Neural Network and compared with statistical model algorithm with four kinds of gestures. The recognition result of method using convolutional neural network is better than other algorithms in EPS gesture signal recognition.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.6
• /
• pp.505-514
• /
• 2020

Since a large amount of training data is typically needed to train Deep Neural Networks (DNNs), a parallel training approach is required to train the DNNs. The Stochastic Gradient Descent (SGD) algorithm is one of the most widely used methods to train the DNNs. However, since the SGD is an inherently sequential process, it requires some sort of approximation schemes to parallelize the SGD algorithm. In this paper, we review various efforts on parallelizing the SGD algorithm, and analyze the computational overhead, communication overhead, and the effects of the approximations.

• KIPS Transactions on Computer and Communication Systems
• /
• v.10 no.3
• /
• pp.71-80
• /
• 2021

The selection of an appropriate neural network algorithm is an important step for accurate data prediction in machine learning. Many algorithms based on basic artificial neural networks have been devised to efficiently predict future data. These networks include deep neural networks (DNNs), recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and gated recurrent unit (GRU) neural networks. Developers face difficulties when choosing among these networks because sufficient information on their performance is unavailable. To alleviate this difficulty, we evaluated the performance of each algorithm by comparing their errors and processing times. Each neural network model was trained using a tax dataset, and the trained model was used for data prediction to compare accuracies among the various algorithms. Furthermore, the effects of activation functions and various optimizers on the performance of the models were analyzed The experimental results show that the GRU and LSTM algorithms yields the lowest prediction error with an average RMSE of 0.12 and an average R2 score of 0.78 and 0.75 respectively, and the basic DNN model achieves the lowest processing time but highest average RMSE of 0.163. Furthermore, the Adam optimizer yields the best performance (with DNN, GRU, and LSTM) in terms of error and the worst performance in terms of processing time. The findings of this study are thus expected to be useful for scientists and developers.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2016.04a
• /
• pp.754-757
• /
• 2016

기존 번호판 인식은 직접 특징 추출 알고리즘을 개발하여 완전 연결 신경망으로 특징을 분류하는 방법이 보편적이다. 본 연구는 전처리 과정에서 번호판 후보군 검출 및 세그먼테이션을 수행하고 특징 추출 없이 미리 학습된 심층 컨볼루션 신경망을 통해 문자를 분류하는 방법을 제안한다. 직접 수집한 2,900장의 번호판 데이터베이스를 이용하여 훈련 집합 및 검증 집합을 구성하였다. 훈련 집합과 검증 집합에 대해 실험한 결과 번호판 후보군 검출률은 97%를 얻을 수 있었고, 이에 대한 인식률은 95%를 얻었다.