基于高效通道注意力网络的人脸表情识别

Facial expression recognition based on high efficientchannel attention network

为了解决现有的人脸表情识别特征提取易受背景及个体因素影响,类内差距大,类间相似度高及实时性较差等问题,提出了一种高效通道注意力网络的轻量级表情识别方法。基于深度可分离卷积改进线性瓶颈结构减少网络复杂性和防止过拟合;通过设计高效注意力模块将特征图的深度与空间信息结合,更着重于重要特征提取,并采用联合损失函数减少相同表情的类内特征差异,扩大不同表情类间特征间距,使网络具有更好的特征判别效果。所提方法在FER-2013与CK+数据集的识别率达到73.3%与97.9%,对比当前诸多较新的方法具有更好的识别性能。

In order to solve the problems that the existing facial expression recognition feature extraction is susceptible to be affected by background and individual factors,gaps within classes are large,similarity between classes are high,and real-time performance is poor,a lightweight facial expression recognition method based on attention network with efficient channels is proposed.This method is based on the depth separable convolution to improve the linear bottleneck structure to reduce network complexity and prevent over-fitting.The depth of the feature map is combined with spatial information by designing efficient attention module,focusing more on the extraction of important features,and adopting a joint loss function to reduce the feature difference within the same expression category,expand the feature spacing between different expression categories,so that the network has better feature discrimination effect.The recognition rate of this method in the FER-2013 and CK+datasets reacheds 73.3%and 97.9%,which has better recognition performance than many current newer methods.