Multi-view Facial Expression Recognition And Pose Estimation Based On Compound Cosine Enhancement Loss

With the development of pattern recognition and artificial intelligence,multi-view facial expression recognition and pose estimation have become the focus of research in the field of computer vision,which have many applications in different areas inside,such as humancomputer interaction,safe driving,clinical medicine,video surveillance.In recent years,facial expression recognition(FER)has achieved the state-of-the-art recognition accuracy for the frontal face images of publicly available databases,which has been reported to be incapable of addressing the complex environments unrelated to facial expressions.Therefore,it is of great value to build a multi-view expression recognition model with good generalization and strong robustness under complex backgrounds.Facial posture plays an important role in the facial expression recognition,which is also great significance to human-computer interaction.At present,the model for multi-view FER is in a step by step fashion.The structure has many shortcomings,such as large parameters,cumbersome steps and poor practicability.In this paper,A multi-view facial expression recognition and pose estimation network with composite cosine enhancement loss is proposed,which can achieve simultaneous recognition of facial expressions and poses and solve the problems of poor accuracy and large step model in complex environments.The research of this paper is multi-view facial expression recognition and pose estimation.First,a deep network framework is built for multi-class recognition in the research.second,we adopted convolutional neural network(CNN)and the existing high-performance deep network structure to achieve multi-view expression recognition and pose estimation(Convolutional Neural Network for Multi-view Facial Expression Recognition and Pose Estimation,MEPCnet),which is in an end-to-end training fashion.Then,the extracted network features are described by the cosine loss in angle space,the performance of which prior of softmax loss.Cosine enhancement loss add the inter-class angle margin parameter and the hypersphere scaling factor to train a model classifier with high intra-class tightness and low intra-class difference,which can further improve the multi-class recognition accuracy.Furthermore,we explain the geometrical meaning of the improved cosine enhancement loss(Cos Eh)through a two-class visualization example and situation of multi-class.Finally,the method of the adaptively adjustable learning is used to optimize the network,which make the loss change more gradual and improve model performance.We conduct experiment on different databases to analyze the multi-classification network model,the results show that the recognition accuracy of the model in the CK + database is as high as 99.06%;the expression recognition rate in the BU-3DFE and Multi-PIE databases are respectively 87.98% and 85.77%,which the pose estimation rates are 99.07% and 85.95% in.The method in this paper has very good performance that compared with the existing methods,which verifying the feature validity and model generalization.Therefore,the model in this paper has some advantages especially in multi-view complex faces,such as novel structure,good performance for complex factors and realizing the simultaneous recognition of facial expressions and gestures.