摘要
针对脑电图(EEG)动态复杂性和解码难度限制情绪识别的精度和鲁棒性的问题,提出一种新的情绪分类模型MADBNet。首先,多尺度分组卷积用于捕捉不同层次的情绪特征;随后,通过分层残差聚合多尺度特征,并穿插轴向通道空间注意力捕获通道相关性和空间依赖性;最后,通过双分支维度分裂特征处理的注意力机制增强局部与全局关联,实现EEG时空频特征的融合。在DEAP数据集上的实验结果表明,该模型在精度和稳定性显示出优越的性能。
Electroencephalograph(EEG)signals exhibit strong dynamic complexity and are inherently difficult to decode,which limits the accuracy and robustness of emotion recognition.To address these challenges,a novel emotion classification model named MADBNet is proposed for efficient EEG representation learning.This model employs multi-scale grouped convolution to extract hierarchical emotional features,and then adopts hierarchical residual aggregation to integrate these features,complemented by axial channel-spatial attention to effectively capture inter-channel relationships and spatial dependencies.Furthermore,a dual-branch dimension-split attention mechanism is introduced to strength local and global feature interactions,thereby realizing the comprehensive fusion of EEG spatiotemporal-frequency characteristics.Experimental results on the DEAP dataset demonstrate that this model achieves superior performance in both accuracy and stability.
作者
李杰
何文雪
王述畅
杨帮华
LI Jie;HE Wenxue;WANG Shuchang;YANG Banghua(School of Automation,Qingdao University,Qingdao 266071,China;School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200444,China)
出处
《中国医学物理学杂志》
2026年第2期220-228,共9页
Chinese Journal of Medical Physics
基金
国家自然科学基金(62376149)。
关键词
情绪识别
人机交互
分层残差聚合
轴向通道空间注意力
双分支维度分裂
emotion recognition
human-computer interaction
hierarchical residual aggregation
axial channel-spatial attention
dual-branch dimension-split
