Research On Dynamic Expression Recognition Algorithm Based On Feature Fusion

Empowering computer the intelligence of human for accurate facial expression recognition,has a significant practical value.However,facial expression recognition is susceptible to uncontrolled environmental factors such as lighting,gestures,and occlusions.At the same time,the small sample capacity of the existing expression dataset,the obvious inter-class similarity and intra-class difference in the sample space both make recognition in trouble.This paper analyzes the expression recognition task in depth from the feature level,aiming to obtain more robust expression features by extracting and fusing features of different properties,thereby significantly improving the recognition effect.The research content and main contributions of this article include:(1)It is difficult to distinguish between similar and intra-class difference samples for single expression features,which is a problem of misclassification.This paper proposes a dynamic expression recognition algorithm based on first and second order spatial feature fusion.The algorithm uses a covariance pooling structure to map facial feature maps to the Riemannian manifold space,and encodes the second-order space features of the expression.This feature exhibits a higher sensitivity to facial organ distortions in the Riemannian manifold.In order to overcome the loss of information in the original space during space mapping,this paper fuses the first and second order spatial features of convolutional layer coding to effectively enrich the information expression of face features in Riemann space,and finally input the fused features into recurrent neural network.The network captures the dynamic change information in time series,realizes dynamic expression recognition and corrects misclassified samples.(2)Due to homologous features are prone to information redundancy during fusion,this paper analyzes the expressive characteristics of the first and second order spatial features,geometric features,and temporal features and the independence and correlation between each other to improve the object of feature fusion.A dynamic expression recognition algorithm based on spatial-geometric feature fusion is proposed.It uses expression sequence pictures and face key points to encode more comprehensive features,suppress interference from complex environmental factors,and further improve recognition accuracy.Meanwhile,the problem that the small sample data set causes the model to be easily over-fitted and the feature discrimination is weak.This paper proposes a step-by-step training method based on transfer learning,which transfers face recognition knowledge to expression recognition tasks in the form of parameter transfer,supplements expression features in the data domain,and improves the training efficiency of the model With generalization performance.The experimental results show that the recognition algorithm based on the fusion of first and second order spatial features proposed in this paper can effectively improve the recognition accuracy of the model,especially the fear and happy expressions have prominent effects.The dynamic expression recognition algorithm model based on spatial-geometry feature fusion proposed in this paper is equivalent to or better than most of the latest methods in recognition accuracy and training speed.