A spatial-frequency-temporal 3D convolutional neural network for motor imagery EEG signal classification

• Published in: Signal, image and video processing Vol. 15; no. 8; pp. 1797 - 1804
• Main Authors: Miao, Minmin, Hu, Wenjun, Zhang, Wenbin
• Format: Journal Article
• Language: English
• Published: London Springer London 01.11.2021; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial neural networks; Classification; Competition; Computer Imaging; Computer Science; Datasets; Electroencephalography; Feature extraction; Human-computer interface; Image classification; Image Processing and Computer Vision; Machine learning; Multimedia Information Systems; Original Paper; Pattern Recognition and Graphics; Signal classification; Signal,Image and Speech Processing; Three dimensional models; Vision; Spatial frequency temporal (SFT); Electroencephalogram (EEG); 3D convolutional neural network (3D CNN); Brain–computer interface (BCI)
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-021-01924-3

Motor imagery (MI) EEG signal classification is a critical issue for brain–computer interface (BCI) systems. In traditional MI EEG machine learning algorithms, feature extraction and classification often have different objective functions, thus resulting in information loss. To solve this problem, a novel spatial-frequency-temporal (SFT) 3D CNN model is proposed. Specifically, the energies of EEG signals located in multiple local SFT ranges are extracted to obtain a novel 3D MI EEG feature representation, and a novel 3D CNN model is designed to simultaneously learn the complex MI EEG features in the entire SFT domains and carry out classification. An extensive experimental study is implemented on two public EEG datasets to evaluate the effectiveness of our method. For BCI Competition III Dataset IVa, the average accuracy rate of five subjects obtained by the proposed method reaches 86.6% and yields 4.1% improvement over the state-of-the-art filter band common spatial pattern (FBCSP) method. For BCI Competition III dataset IIIa, by achieving an average accuracy rate of 91.85%, the proposed method outperforms the state-of-the-art dictionary pair learning (DPL) method by 4.44%.