全局卷积和双注意力机制在动态表情识别中的应用
Application of global convolution and dual attention mechanism in dynamic expression recognition
梁成旭 1董建设 1李金良1
作者信息
- 1. 天津职业技术师范大学信息技术工程学院,天津 300222
- 折叠
摘要
静态面部表情识别无法捕捉表情的动态变化,识别过程中丢失了表情的连续性特征,同时现实情境中存在的表情强度差异给识别带来不利影响.针对以上问题,提出一种全局卷积双注意力机制的强感知学习神经网络模型,对动态表情序列进行识别.该模型通过全局卷积双注意力块重新缩放特征映射通道,引入全局卷积双注意力块(squeeze and excitation-global convolution attention,GCSA)增强残差网络的学习能力,提高表情识别的准确率.在训练的过程中,引入强感知交叉熵损失函数(cross entropy-auxiliary-intensity aware loss,CAI)来处理视频序列中存在不同感知强度帧的问题.通过在公开数据集DFEW和FERV39K上进行实验,从客观指标和主观视觉效果上与经典方法进行比较.结果表明,所提方法在多个性能指标上均优于对比方法,证明了方法的有效性.
Abstract
Static facial expression recognition fails to capture the dynamic changes of expressions,resulting in the loss of continuity features during the recognition process.Additionally,discrepancies in expression intensity in real-world contexts adversely affect the recognition accuracy.To address these challenges,this paper proposes a robust perceptual neural net-work model that leverages global convolution and a dual attention mechanism for recognizing dynamic expression sequences.This model incorporates a global convolutional dual attention block(squeeze and excitation-global convolution attention,GCSA)to rescale the channels of feature maps,thus enhancing the learning capabilities of the residual network and im-proving the accuracy of expression recognition.To tackle the issue of varying perceptual intensities across frames in video sequences,a strong perceptual cross-entropy loss function(cross-entropy-auxiliary-intensity aware loss,CAI)is in-troduced during the training process.Experiments conducted on the publicly available datasets DFEW and FERV39K demonstrate that the proposed method outperforms classical approaches in multiple performance metrics,validating the ef-fectiveness of this approach through both objective measures and subjective visual assessments.
关键词
表情识别/动态表情序列/注意力机制/表情强度Key words
expression recognition/dynamic expression sequence/attention mechanism/expression intensity引用本文复制引用
基金项目
天津市自然科学基金资助项目(22JCYBJC00470)
出版年
2024
NETL
NSTL
万方数据