Facial Action Unit Recognition And Its Application In Expression Transfer System

The study of human facial behavior is one of the important research topics in the field of computer vision and psychology and artificial intelligence.Compared with the expression categories of facial expressions,facial action units can more fully express human facial behavior.The current facial motion unit recognition research is mainly based on traditional SVM classification and the popular deep learning methods.However,these methods regard the human face as a whole,which leads to the interference of unrelated parts in the recognition of facial action units.To this end,this paper proposes a facial action unit recognition method based on block deepening network and a facial action unit recognition method based on RPN network.The main research contents of the paper include:(1)A facial action unit recognition method based on block deepening network is proposed.First,the face detection method is used to detect faces in the image,and then the images containing the faces are normalized to remove parts that are not related to facial behavior.Secondly,block-based convolution is used to deepen the standardized face of the network plane to perform block-based convolution,so as to map different AUs to different blocks,and avoid the interference of AU recognition by some blocks of unrelated planes.The network combines the network structure of the currently popular target detection network,and furthermore,enables the network to extract facial features more effectively;finally,the classification and regression of the extracted Feature Map obtains the results of AU recognition.The experimental results show that the proposed facial motion unit recognition method based on block deepening network is superior to the traditional SVM-based AU recognition and the more popular fully-supervised AU recognition method.(2)A facial action unit recognition method based on RPN network is proposed.The macro idea of this method is to reduce the influence of image-independent factors on facial action unit recognition.Different from the facial action unit recognition method of block deepening network,this method mainly improves the end of the network.First,the input image is subjected to face positioning and cropping,and the cropped image is input to a feature extraction network for feature extraction.Then use the RPN network to locate the key position of the face(parts strongly related to AU),and the feature map of the located key position for basic facial expression recognition.The core of facial action unit recognition is to identify the feature map corresponding to the key position to the associated AU,and the expression verification is added in the recognition process.Experiments show that our method achieves better results in both facial expression recognition and facial action unit recognition.(3)Based on the above work,a facial expression migration system is implemented.The system mainly realizes real-time migration of facial expressions,definition of expression files and extraction of expression files in videos.The system mainly adopts the facial motion unit recognition method based on the block deepening network proposed in this paper,and uses a variety of classic face detection methods.