Research On FER Based On Domain Information Loss And Attention Dynamic Weighted Training

With the rapid development of artificial intelligence technology in recent years,facial expression recognition has appeared in a wide range of application scenarios in the fields of intelligent security,polygraph detection,smart medical care,and Internet education.Due to the complexity of facial expressions,it is difficult to obtain the feature distribution of facial expressions by artificially designing and selecting features.Therefore,the methods of automatically extracting facial expression features based on deep convolutional neural network are played an important role in facial expression recognition.The facial expression recognition task has always been a challenging problem in the field of computer vision.Due to the interference of individual emotion expression and the uncontrolled environment,it is difficult to exactly define the feature space for each expression class.Mainly,it will lead to the problem of directional misclassification between facial expression classes.In order to alleviate this problem,we propose the domain information loss function to guide the feature learning of networks by adding the priori domain information into objective function.At the same time,it is a difficult task to classify images with multiple class labels using only a small number of labeled examples.In this paper,we propose a data augmentation networks named Unet-StarGAN generative adversarial networks to randomly generate the expression images with the same distribution of facial expression datasets.In order to avoid gradient disappearance problem and effectively capture such distortions in regional facial features,we add the covariance pooling layer and the residual block into the deep convolutional neural network.In addition,the objective function of facial expression network include multiple loss functions with weighting coefficients.Therefore,in this paper we propose a attention dynamic weighted training method,which divides the whole network training process into three different stages.Our network can learn facial features more efficiently and reasonably throughout the training process by dynamically adjusting the weighting coefficients at each training stage.It can be seen from the experimental results that our domain information loss and attention dynamic weighted training method significantly not only improves the performance of facial expression recognition but also alleviates the problem of directional misclassification between facial expression classes.Extensive experiments on various recognition benchmarks like RAFDB,FER-2013 datasets verify the effectiveness of our proposal.In addition,the experimental results show that our proposed method outperforms the state-of-art recognition accuracy on the RAF-DB datasets and achieves competitive recognition performance on the FER-2013 datasets.