Dynamic Expression Recognition Based On Multi-feature Deep Fusion

With the development of artificial intelligence,facial expression recognition in the field of machine vision has received more and more attention,widely used in the fields of automobile driving,medical treatment,monitoring,distance education and human-computer interaction.Because of the facial expression is a continuous process,the expression based on the moving image sequence contains spatial information and time domain information,which can obtain a better recognition rate than the static image,and thus becomes a hot issue in expression recognition research.Since the existing method extracts only the features of the shallow layer,the deep layer or the simple fusion of the two,it is difficult to obtain the local and global expression information at the same time,thus causing the expression recognition to fail to meet the actual needs.Aiming at the above problems,this paper studies the dynamic facial expression feature extraction method and proposes a multi-feature deep fusion method to improve the accuracy of expression recognition.The innovations and main work of this paper include:A triangle geometric semantic feature method is proposed to extract shallow local features.Firstly,the peak frame detection method is used to extract the peak frame of the input sequence,then the feature point is used to calibrate the expression triangle to extract the triangle geometric feature.the triangle geometric features are semantically described,the semantic feature description set is constructed,and then the feature intensity level discriminant rules are formulated.Finally,the triangle geometric features are analyzed to obtain the optimal triangle combination and the feature strength is obtained.After the level is judged,the triangle geometric semantic feature is obtained.The triangle geometric semantic feature effectively reduces the time complexity while highlighting the differences of expressions.A dynamic facial expression recognition method with multi-feature depth fusion is proposed to extract the fusion features of recognition expressions.The time series interpolation frame model is used to preprocess the moving image sequence to obtain the same image sequence.Then the 3D-Res Net network is used to extract the global depth feature of the image sequence,After the two-layer 2D convolution operation is performed on the obtained triangle geometric and depth features,the feature fusion is performed through two-channel parallel connection.Finally,the convolutional neural network(CNN)is used to convolve the fusion features and learn.The relationship between the global feature and the local feature obtains more discriminative features and is input into the Softmax classifier for dynamic expression recognition.In this paper,the experimental verification is carried out on the CK+ and MMI expression database.Compared with the mainstream dynamic expression recognition method,the effectiveness of the proposed method is verified.