Research On Facial Expression Recognition Based On Subspace Embedding And Transfer Learning

With the development of science and information technology,facial expression recognition technology plays an important role in behavioral science,medical science,computer science and many other fields.However,there are still many problems waiting to be solved for facial expression recognition due to many factors such as the high-dimensional data of facial expression image,light,posture,and the diversity of facial expressions in practical application.Based on subspace embedding and transfer learning algorithm,this paper aims to study facial expression recognition.The main research contents are as follows:(1)Aiming at the problem that the traditional LBP features cannot describe the details of expression images meticulously,and at the same time,in order to maintain the robustness of the interference factors such as light and angle,the block LBP features are first extracted from the expression images.Then in the process of dimensionality reduction based on graph embedding technology,the intrinsic adjacency graph is constructed by using the within-class scatter constraints to enhance the reconstruction relationship between given samples and non-neighbor samples from the same class.The inter-class scatter constraint is used to construct the penalty adjacency graph to reduce the impact of reconstruction between given samples and pseudo-neighbor samples of different classes.In addition,in the process of low-dimensional projection,global distribution constraints are introduced into the projection objective function to find the optimal subspace,achieving the purpose of reducing within-class distance and increasing inter-class distance while reducing dimensions.(2)Aiming at the problem,caused by distribution mismatching,in cross-domain expression recognition of expression image data,a learning and recognition method of expression subspace guided by single data is proposed.Joint low-rank and sparse constraints are imposed on the reconstruction coefficient matrix to preserve the global and local structures of facial expression data.In order to expand the margins among different classes and provide more freedom to decrease the discrepancy,a flexible linear classifier(projection)can be acquired by learning anon-negative label relaxation matrixthat allows the strict binary label matrix to relax into a slack variable matrix.Transfer both the source and target data to a common subspace by using a transformation matrix so that the discrepancy of the source and target domains can be reduced.(3)An expression recognition algorithm based on terms guide subspace learning method is proposed,which means that a two-stage learning strategy with teacher teaching and student feedbacks uses three guidance terms to learn an invariant,discriminative,and domain-agnostic subspace.Firstly,the subspace-guided term moves the source nearer to the target subspace to decrease the discrepancy between domains.Secondly,the data-guided term applies the coupled projections to map both domains to a common subspace,in which every target sample can be represented by the source samples with low-rank coefficient matrix which can maintain the global structure of facial expression data.Thirdly,to improve the discrimination of the subspaces,the label-guided term is constructed to predict based on source labels and pseudo-target labels.Furthermore,a label relaxation matrix is introduced to improve the model tolerance to label noise.