基于通道注意力的人脸表情识别

Expression recognition based on channel attention

目的:解决通道注意力提取过程中各个通道信息利用不充分、交互性不足的问题。方法:提出一种基于SENet改进的通道注意力模型。本模型利用Inverted Bottleneck提取更加完整的通道信息。将GELU函数引入表情识别,以改善网络升维操作带来的过度融合问题。同时利用信息熵判断不同卷积核生成特征图的重要程度,为网络引入更多的归纳偏置。结果:在CK+和Oulu-CASIA库人脸表情数据集上的正确率分别达到95.92%和91.21%。结论:本方法能够更加充分地利用各通道特征,在有效提升准确率的同时具有较好的泛化能力。

Objective:To improve the current problem of insufficient utilization of information and insufficient interactivity in channel attention extraction.Methods:An improved channel attention model based on SENet was proposed.This model utilizes Inverted Bottleneck to extract more complete channel information.Introduce the GELU function into facial expression recognition to improve the over fusion problem caused by network dimensionality operations.At the same time,information entropy is used to judge the importance of different convolution kernels to generate feature maps,so as to introduce more inductive bias into the network.Results:The accuracy rates on the CK+and Oulu-CASIA facial expression datasets reached 95.92%and 91.21%,respectively.Conclusion:The method proposed in this article can more fully utilize the features of each channel,effectively improving accuracy while possessing good generalization ability.