摘要
人脸表情的表达受到全局特征和局部特征的共同作用,为了高效融合并利用这些特征,提出了一种基于双分支和多尺度注意力机制的表情识别方法 .方法包括双分支特征提取、多级别特征融合和多级别特征注意力三个模块.在双分支特征提取模块,将原图像和经过小波变换得到的高频图像组成双通道分别经过主干网络提取特征,其中原图像提供了全局特征,高频图像提供了像素级别的局部特征.在多级别特征融合模块,将双通道主干网络不同阶段提取的多尺度特征图进行融合,进一步提取局部细节和全局信息.在多级别特征注意力模块,将融合后的多尺度特征分块,统一各块的维度后送入自注意力模块并获得预测的类别.该模型在CK+和RafDB两个表情库上进行实验,通过和近年来的优秀模型做对比,验证了方法的有效性.
The facial expression is influenced by both global and local features.In order to efficiently integrate and utilize these features,a facial expression recognition method based on dual branch and multi-scale attention mechanism is proposed.The method includes three modules:dual branch feature extraction,multi-level feature fusion,and multi-level feature attention.In the dual branch feature extraction module,the original image and the high-frequency image obtained through wavelet transform are combined into two channels by passing through the backbone network to extract features.The original image provides global features,while the high-frequency image provides pixel level local features.In the multi-level feature fusion module,the multi-scale feature maps extracted at different stages of the dual channel backbone network are fused to further extract local details and global information.In the multi level feature attention module,the fused multi-scale features are divided into blocks,and the dimensions of each block are unified before being fed into the self attention module to obtain the predicted category.This model has been extensively tested on two expression libraries,CK+ and RafDB,and its effectiveness has been verified by comparing it with excellent models in recent years.
作者
李堃
盛家乐
李锐
王倩倩
LI Kun;SHENG Jiale;LI Rui;WANG Qianqian(Anhui Science and Technology University,Bengbu 233000,China)
出处
《通化师范学院学报》
2025年第8期34-42,共9页
Journal of Tonghua Normal University
基金
安徽省教育厅重点项目(2022AH051638)
安徽省高等学校自然科学研究项目(2023AH051867)
安徽省高校自然科学研究重点项目(2024AH050337)
安徽科技学院引进人才项目(200192-引进人才项目XWYJ202004)。
关键词
双分支
小波变换
多尺度
注意力机制
dual branch
wavelet transform
multi-scale
attention mechanism
