Yunyuan Gao1, Xiangkun Wang2, Thomas Potter3, Jianhai Zhang4, Yingchun Zhang5. 1. Intelligent Control and Robotics Institute, College of Automation, Hangzhou Dianzi University, Hangzhou, China; Key Laboratory of Brain Machine Collaborative Intelligence of Zhejiang Province, Hangzhou, China. Electronic address: gyy@hdu.edu.cn. 2. Intelligent Control and Robotics Institute, College of Automation, Hangzhou Dianzi University, Hangzhou, China. 3. Department of Biomedical Engineering, University of Houston, Houston, USA. 4. Key Laboratory of Brain Machine Collaborative Intelligence of Zhejiang Province, Hangzhou, China. 5. Department of Biomedical Engineering, University of Houston, Houston, USA. Electronic address: yzhang94@uh.edu.
Abstract
BACKGROUND: Emotion recognition has been studied for decades, but the classification accuracy needs to be improved. NEW METHOD: In this study, a novel emotional classification approach is proposed by combining the Histogram of Oriented Gradient (HOG) method with the Granger Causality (GC) or Transfer Entropy (TE) methods. HOG extracts local valid information from the GC/TE relationship matrices and then the Support Vector Machine (SVM) is employed to classify the emotional states of stress and calm. RESULTS: Compared with previous studies, the classification accuracy has been greatly improved. The results of this study show that the classification based on GC or TE with HOG offers an average accuracy 88.93 % and 95.21 %, respectively. The achieved accuracy is about 12 % higher than that achieved without using HOG feature extraction. COMPARISON WITH EXISTING METHOD(S): Numerous efforts have been devoted to classify emotional states by extracting EEG characteristics on a single channel basis, the method developed in this study utilizes information interaction between brain channels as a feature to classify emotional states. Furthermore, this study combines HOG and relation matrices for the first time and uses image processing to extract EEG features. CONCLUSION: Our results demonstrate the feasibility of combining TE with HOG for emotion recognition with improved classification accuracy by taking advantage of both network and gradient features. More specific features can be selected to improve classification accuracy by taking advantage of information exchanges between EEG channels directly or the extracted property of the relationship matrix based on information interactions.
BACKGROUND: Emotion recognition has been studied for decades, but the classification accuracy needs to be improved. NEW METHOD: In this study, a novel emotional classification approach is proposed by combining the Histogram of Oriented Gradient (HOG) method with the Granger Causality (GC) or Transfer Entropy (TE) methods. HOG extracts local valid information from the GC/TE relationship matrices and then the Support Vector Machine (SVM) is employed to classify the emotional states of stress and calm. RESULTS: Compared with previous studies, the classification accuracy has been greatly improved. The results of this study show that the classification based on GC or TE with HOG offers an average accuracy 88.93 % and 95.21 %, respectively. The achieved accuracy is about 12 % higher than that achieved without using HOG feature extraction. COMPARISON WITH EXISTING METHOD(S): Numerous efforts have been devoted to classify emotional states by extracting EEG characteristics on a single channel basis, the method developed in this study utilizes information interaction between brain channels as a feature to classify emotional states. Furthermore, this study combines HOG and relation matrices for the first time and uses image processing to extract EEG features. CONCLUSION: Our results demonstrate the feasibility of combining TE with HOG for emotion recognition with improved classification accuracy by taking advantage of both network and gradient features. More specific features can be selected to improve classification accuracy by taking advantage of information exchanges between EEG channels directly or the extracted property of the relationship matrix based on information interactions.
Authors: Behrad Soleimani; Proloy Das; I M Dushyanthi Karunathilake; Stefanie E Kuchinsky; Jonathan Z Simon; Behtash Babadi Journal: Neuroimage Date: 2022-07-21 Impact factor: 7.400