重庆邮电大学学报（自然科学版）2023，Vol.35Issue(6)：1107-1116,10.DOI:10.3979/j.issn.1673-825X.202209030228
自注意力加权半脸的人脸表情识别算法
Facial expression recognition based on self-attention weighted half-face
摘要
Abstract
Aiming at the problems of head deflection and uncertain samples in the dataset,a self-attention weighted half-face facial expression recognition method is proposed.First,the face area and the left and right half-face areas are searched in the input image and sent to the multi-feature fusion backbone network with a specific structure to extract features.Then,the self-attention weight module assigns appropriate weights to the left and right half-faces as the auxiliary clues of expres-sion prediction.Finally,the parallel-line prediction fusion module combines feature map and weight information to make ex-pression predictions.In addition,this paper designs an adaptive relabeling module to help the model locate uncertain sam-ples in the dataset and assign them appropriate pseudo-labels.Experiments show that the recognition accuracy of 90.76％,91.08％,and 98.66％are obtained on RAF-DB,FERPlus,and RaFD datasets,respectively.关键词
人脸表情识别/多特征融合/注意力机制/自适应方法Key words
facial expression recognition/multiple feature fusion/attention mechanism/adaptive method分类
计算机与自动化引用本文复制引用
汪榕涛,黎勇,刘锐,刘泽圣..自注意力加权半脸的人脸表情识别算法[J].重庆邮电大学学报（自然科学版）,2023,35(6):1107-1116,10.基金项目
国家自然科学基金项目(61771081)The National Natural Science Foundation of China(61771081) （61771081）
