EMD를 이용한 EEG 신호 기반의 감정인식

이다빛(David Lee),이상국(Sang-Goog Lee) 한국정보기술학회 2014 한국정보기술학회논문지 Vol.12 No.6

Human-Computer Interface(HCI) needs recognizing the users emotion in order to provide good user satisfaction. In this paper, we propose a method of recognizing user`s emotions from Electroencephalogram(EEG) signal using Empirical Mode Decomposition(EMD) that effectively decomposes non-linear and non-stationary signals. The emotion recognition consists of the feature extraction phase and the classification phase. In the feature extraction process, the EMD was used to generate Implicit Mode Functions(IMFs) from the EEG signal. Then the Fast Fourier Transform(FFT) was used to identify frequency band of generated IMFs. The power spectral density(PSD) value of the IMF that included alpha(8-12Hz) and beta(13-30Hz) was used as a feature. In the classification process, Support Vector Machine(SVM) was used to classify extracted features. Then 5-fold cross validation was utilized to evaluate the classification accuracy of emotion recognition. The proposed method provided classification accuracy of 80.25% and 74.56% for valence and arousal.

Motor Imagery EEG Classification Method using EMD and FFT

이다빛(David Lee),이희재(Hee-Jae Lee),이상국(Sang-Goog Lee) Korean Institute of Information Scientists and Eng 2014 정보과학회논문지 Vol.41 No.12

상호 정보와 희소 학습을 이용하여 동작 상상 뇌파의 분류 정확도 향상

이다빛(David Lee),박상훈(Sang-Hoon Park),이희재(Hee-Jae Lee),이상국(Sang-Goog Lee) 한국정보기술학회 2017 한국정보기술학회논문지 Vol.15 No.8

Brain-computer interface (BCI) is a technology that directly connects a human brain and a computer to control and manipulate the computer through electroencephalogram. The common spatial pattern (CSP) is the most famous method for extracting discriminative features from BCI based on motor imagery electroencephalogram. However, CSP is sensitive to the operational frequency band, and this operational frequency band is subject-specific. In this paper, we propose a method to extract subject-specific features based on mutual information and sparse learning. The proposed method divides the 4~40Hz band into 17 sub-bands using a fifth order Butterworth filter. Then, CSP is applied to each sub-band to extract CSP features. Thereafter, mutual information is used to select the four most discriminating feature sets from the CSP features and sparse learning was used to eliminate redundancy in the four pairs of features selected. Finally, support vector machine was used for learning and classification. The performance of the proposed method was evaluated using formally available BCI Competition III and IV. The proposed method showed higher average classification accuracy than the results of CSP, FBCSP and SFBCSP.

컨볼루션 신경망을 이용한 동작 상상 뇌파 분류

이다빛(David Lee),박상훈(Sang-Hoon Park),이희재(Hee-Jae Lee),이상국(Sang-Goog Lee) 한국정보기술학회 2017 한국정보기술학회논문지 Vol.15 No.6

Brain-computer interface (BCI) is a technology that can be used as augmentative and alternative communication (AAC) for people such as the elderly or the disabled who are restricted or impaired in physical function. In order for BCI to be used as AAC, it is important to select appropriate feature extraction and classification methods because the electroencephalogram (EEG) signal is non-linear and non-stationary. This study proposes a feature extraction and classification method of motor imagery EEG using convolutional neural network (CNN). The CNN, most commonly used in the field of images, uses a large number of training data to avoid the problem of overfitting. If the amount of training data is small, the CNN cause overfitting problems. Therefore, in this study, the CNN suitable with small amount of training data was designed for motor imagery based BCI, and then the motion imaginary EEG was learned and classified. The performance of the proposed method is shown to be about 3.8~4.5% in terms of average accuracy through comparison with existing machine learning methods.

웨이블릿 변환과 주성분 분석 기반의 결합 특징 벡터를 이용한 동작 상상 EEG 분류

이다빛(David Lee),박상훈(Sang-Hoon Park),이희재(Hee-Jae Lee),이상국(Sang-Goog Lee) 한국정보기술학회 2017 한국정보기술학회논문지 Vol.15 No.4

The brain-computer interface is a communication channel that controls and manipulates mechanism by transmitting human thoughts from a human to a mechanism. To do this, it is necessary to extract appropriate electroencephalogram (EEG) features. In this paper, we propose a combined feature vectors based on wavelet transform and principal component analysis to extract features of EEG signals. The proposed method consists of three steps: In the first step, the wavelet transform is applied to extract feature vectors of the motor imagery EEG. In the second step, the principal component analysis was used to reduce the dimensionality of the feature vectors. In the third step, the nonlinear support vector machines was applied to classify left or right hand motor imagery EEG. The performance of the proposed method was evaluated in terms of accuracy through three different motor imagery EEG datasets. The results of this study showed that the proposed method achieved higher accuracy in the motor imagery EEG classification. The results of this study show that the proposed method improve the classification accuracy of some subjects by 0.4~4.3% compared to existing single feature methods, thus achieves a high classification accuracy of 82.8%.

GMM과 SVM을 이용한 움직임 상상 뇌파 분류에 관한 연구

이다빛(David Lee),이상국(Sang-Goog Lee) 한국정보기술학회 2013 한국정보기술학회논문지 Vol.11 No.7

The Brain-Computer Interface (BCI) use the Electroencephalogram (EEG) as a method to replace existing interface technologies. Therefore, BCI is the necessary technology not only for general users but also for disabled who can not move the muscle because of the nervous system problem and the senior who are uncomfortable with movement. In this paper, we propose a method which uses the support vector machine (SVM) with the support vector that is generated by the Gaussian mixture model (GMM) to classify movement imagery EEGs. This support vector was robust to noise and overfitting. The EEG classification consists of the feature extraction and the classification process. In the feature extraction process, the wavelet transform was used to decompose the EEG and extract the statistical feature of the wavelet coefficient. In the classification process, the expectation maximization (EM) algorithm was used to obtain the maximum likelihood estimation (MLE) of the distribution of statistical feature of the GMM. Lastly, we utilized the average of the estimated Gaussian distribution to generate the support vector. Using the proposed method, the classification accuracy of movement imagery EEGs was found to be 83.61%.

A Dual Filter-based Channel Selection for Classification of Motor Imagery EEG

이다빛(David Lee),이희재(Hee Jae Lee),박상훈(Snag-Hoon Park),이상국(Sang-Goog Lee) Korean Institute of Information Scientists and Eng 2017 정보과학회논문지 Vol.44 No.9

GMM 기반의 Support Vector 생성을 이용한 움직임 상상 EEG 분류 알고리즘에 관한 연구

이다빛(David Lee),김재호(Jae-Ho Kim),정우혁(Woo-Hyuk Jung),이희재(Hee-Jae Lee),이상국(Sang-Goog Lee) 한국HCI학회 2012 한국HCI학회 학술대회 Vol.2012 No.1

본 논문에서는 노이즈와 overfitting에 강건한 성능을 나타내는 Gaussian Mixture Moel(GMM)로 생성한 Support Vector와 분류 알고리즘인 Support Vector Machine(SVM)을 이용하여 왼손 또는 오른손 움직임 상상 Electroencephalogram(EEG)을 분류하는 방법을 제안한다. EEG 분류는 특징 추출 과정과 분류 과정으로 구성된다. 특징 추출 과정에서는 Wavelet Transform(WT)을 사용하여 EEG를 분해하고 웨이블렛 계수의 통계적 특징을 추출하였다. 분류 과정에서는 가우시안 혼합 모델에 대한 통계적 특징 분포의 최대 우도 추정(Maximum Likelihood Estimation : MLE)을 EM(Expectation Maximization) 알고리즘을 이용하여 얻었다. 그리고 추정된 가우시안 분포의 평균을 가지고 Support Vector를 생성하였다. 제안하는 분류 방법을 이용하여 움직임 상상 EEG에 대해 83.61%의 분류 정확도를 보였다. In this paper, we used a support vector that was generated by a Gaussian mixture model (GMM). This support vector showed robust performance against noise and overfitting, and the support vector machine (SVM) of the classification algorithm suggested a method of classification for left- and right-hand movement imagery electroencephalograms (EEGs). The EEG classification consists of a feature extraction and classification process. In the feature extraction process, a wavelet transform (WT) was used to decompose the EEG and extract the statistical features of the wavelet coefficient. In the classification process, an expectation maximisation (EM) algorithm was used to obtain a maximum likelihood estimation (MLE) of the distribution of statistical features for the GMM. Lastly, we utilised the average of the estimated Gaussian distribution to generate the support vector. Under the proposed method, the classification accuracy of movement imagery EEGs was found to be 83.61%.

EMD를 이용한 EEG 기반 움직임 상상 분류

이다빛(David Lee),김재호(Jae-Ho Kim),정우혁(Woo-Hyuk Jung),이희재(Hee-Jae Lee),이상국(Sang-Goog Lee) 한국HCI학회 2014 한국HCI학회 학술대회 Vol.2014 No.2

Brain-Computer interfaces(BCIs)에서 Electroencephalogram(EEG)의 특징을 추출하는 것은 중요하다. 일반적으로 EEG의 특징 추출 방법으로는 Fast Fourier transform(FFT)과 Wavelet transform(WT)이 많이 사용되었다. 하지만 이러한 방법들은 신호가 linear하고 stationary 하다는 가정 하에 적용되었기 때문에 신호 분해시 신호의 왜곡이 생길 수 있다. 이에 본 논문은 움직임 상상 EEG 분류를 위해 Empirical Mode Decomposition(EMD)과 FFT를 이용하는 특징을 제안했다. 먼저 움직임 상상 EEG에 EMD를 적용하여 Implicit Mode Functions(IMF)를 추출 뒤, 추출된 IMFs에 FFT를 적용하여 해당 IMF의 주파수 성분을 확인하였다. 주파수 성분이 μ 대역을 포함하고 있는 IMF의 표준편차를 특징으로 사용하였다. 추출된 특징을 Support Vector Machine(SVM)의 입력으로 사용하였고 샘플의 검증을 위해 10-fold cross validation을 이용하였다. 제안하는 방법은 움직임 상상 EEG에 대해 84.50%의 분류 정확도를 보여주었다. Feature extraction of Electroencephalogram (EEG) is an important issue in brain-computer interfaces(BCIs). The most commonly used methods for feature extraction from EEGs is Fast Fourier transform(FFT) and Wavelet transform(WT). However, when signal decomposition is carried out , these methods can happens distortion of the signal because it assumes that the signal is linear and stationary. In this paper, we proposed to use Empirical Mode Decomposition(EMD) and FFT to feature for classification of movement imagery EEGs. The EMD was applied to generate Implicit Mode Functions(IMFs) from the movement imagery EEGs. The FFT was then used to identify frequency component of each IMF at generated IMFs. The standard deviation of IMF included mu rhythm was used as feature. In the classification process, we used the extracted feature as input of Support Vector Machine(SVM) and 10-fold cross-validation to verification of sample. Under the proposed method, the classification accuracy of movement imagery EEGs was found to be 84.50%.

EMD와 FFT를 이용한 EEG 기반 감정 인식에 관한 연구

이다빛(David Lee),김재호(Jae-Ho Kim),정우혁(Woo-Hyuk Jung),이희재(Hee-Jae Lee),이상국(Sang-Goog Lee) 한국HCI학회 2013 한국HCI학회 학술대회 Vol.2013 No.1

HCI가 기존의 한계를 넘어 HHI로 가기위해서는 감정을 이해하는 능력이 필요하다. 최근 많은 연구들은 음성, 얼굴 표정, 생체 신호등을 이용하여 감정을 인식하였다. 본 논문은 EMD와 FFT를 이용하여 뇌파에서 사람의 감정을 인식하는 연구를 하였다. 실험 데이터는 사람의 감정을 2차원(Arousal과 Valence) 공간에 분류시킨 DEAP database를 이용하였다. 먼저 감정과 관련된 특징을 추출하기 위해 non-stationary이고 non-linear한 신호인 EEG(Electroencephalogram)에 EMD(Empirical Mode Decomposition)를 적용하여 IMFs(Implicit Mode Functions)를 생성하였다. 그리고 생성된 IMFs에 FFT를 적용하여 해당 IMF의 주요 주파수 성분을 확인하였다. 특징은 주요 주파수 성분이  (8-12 Hz)와  (13-30 Hz)에 포함되는 IMF의 최대 주파수 값을 사용하였다. 추출된 특징을 SVM(Support Vector Machine)의 입력으로 사용하였고 5-fold cross validation을 이용하여 감정 인식의 정확도를 평가하였다. 제안하는 분류 방법을 이용하여 arousal과 valence에 대해 74.56%와 80.25%의 분류 정확도를 보였다. The ability to understand the emotions needs for going to the HHI beyond the HCI. Recently, Many studies have shown that emotion recognition has been studied using speech, facial expressions, and physiological signals. In this paper, we have recognized human emotions from EEG using the EMD and FFT. The DEAP dataset that categorizes various kind of emotions in a two-dimensional space with arousal and valence was used as the experimental data. In the feature extraction process, the EMD(Empirical Mode Decomposition) was used to generate IMFs(Implicit Mode Functions) from the EEG which have non-stationary and non-linear properties. The FFT was used to identify main frequency component of each IMF at generated IMFs. The max frequency value of IMF that included alpha(8-12 Hz) and beta(13-30 Hz) was used as feature. In the classification process, we utilised the extracted feature as input of SVM and 5-fold cross validation to evaluate classification accuracy of emotion recognition. Under the proposed method, the classification accuracy for arousal and valence was 81.56% and 80.25%.