基于深度可分离卷积的表情识别改进方法

High-precision expression recognition method for complex illumination

针对传统表情识别存在相似表情识别精度不高,且深度学习模型参数量巨大问题,提出一种改进的残差网络模型。通过引入深度可分离卷积核,减少了模型的参数量;引入压缩激励模块,改善了模型通道的加权关系;通过将中心损失引入联合算法设计中,提高了相似表情之间的区分度。实验结果表明,识别算法提升了相似表情的区分精度,且较好的控制了模型的参数量。模型在3个公开数据集上的准确率分别达到了97.57%、96.24%、94.09%。

To solve the problems of low accuracy and large number of deep learning model parameters in traditional facial expression recognition,an improved residual network model was proposed.The depth separable convolution kernel was introduced to reduce the number of model parameters,and the compression excitation module was introduced to improve the weighted relationship of model channels.The center loss was introduced into the design joint algorithm to improve the degree of discrimination between similar expressions.The experimental results show that the recognition algorithm improves the discrimination accuracy of similar expressions and controls the number of parameters in the model well.The accuracy of the model on three public data sets is 98.17%,97.22%and 94.09%.