南京邮电大学学报（自然科学版）2025，Vol.45Issue(3)：77-86,10.DOI:10.14132/j.cnki.1673-5439.2025.03.009
基于不对称非局部高效信道注意时空网络面部表情识别
Facial expression recognition based on asymmetric non-local efficient channel attention spatial-temporal network
摘要
Abstract
In order to extract and learn facial expression features from multiple dimensions including the channel domain,time domain,and spatial domain,this paper proposes a novel asymmetric non-local ef-ficient channel attention spatial-temporal network(ANECASN)and applies it to facial expression recog-nition.ANECASN consists of three main modules.The first module is a lightweight asymmetric non-local module,which captures long sequence dependencies and uses a pyramid aggregation optimized feature selection mechanism to obtain better local emotional features.This module improves the low computa-tional efficiency and low emotion recognition rate of deep convolutional operations on long sequences.The second module is an efficient channel attention module,which assigns high weights to channels that are favorable for emotion recognition.It achieves local cross-channel interaction without dimension reduc-tion,reduces model complexity,and enhances the nonlinear expression ability of emotions to achieve performance improvement.The third module is the spatio-temporal LSTM module,which promotes infor-mation interaction between space and time by learning the spatial correlation of emotional features and the temporal correlation of emotional feature sequences.The experiments are conducted on the Multi-modal and RAMAS databases.The results show that ANECASN obtains a recognition rate of 61.54％on the Multimodal database and 42.49％on the RAMAS database,reaching at least a 5％improvement com-pared to the baseline ResNet-50.关键词
人脸表情/不对称非局部高效信道注意时空网络/不对称非局部/高效注意信道/时空LSTMKey words
facial expression/asymmetric non-local efficient channel attention spatial-temporal network(ANECASN)/asymmetric non-local/efficient channel attention/spatio-temporal LSTM分类
计算机与自动化引用本文复制引用
闫静杰,孙雯静,顾晓娜,周晓阳,魏金生..基于不对称非局部高效信道注意时空网络面部表情识别[J].南京邮电大学学报（自然科学版）,2025,45(3):77-86,10.基金项目
国家自然科学基金(61501249)和区块链技术与数据安全工业和信息化部重点实验室开放课题(20242218)资助项目 （61501249）
