Research Of Expression Recognition With Low-rank Cooperative Dictionary Learning Under Unconstrained Environment

With the development of artificial intelligence,facial expression recognition technology has broad application prospects in human-computer interaction,online education,entertainment,medical monitoring,etc.After decades of research,facial expression recognition under constrained conditions has achieved good results.However,in the real world,facial expressions exhibit various attributes due to various factors such as occlusion,posture,and uneven lighting.At the same time,due to data diversification,data imbalance factors also bring great challenges.In this paper,the research on the two problems of occlusion and imbalance of data sets in expression recognition under unconstrained environment is carried out,and expression recognition frameworks are proposed respectively.The specific research content and innovative work are as follows:(1)In the unconstrained environment,facial expression images are often accompanied by various occlusions and other noise information,which does not meet the linear subspace assumption.In addition,due to the insufficient number of training images,certain types of test images may not be formed by the training images on the subspace.In order to solve the above problems,a facial expression recognition framework based on the dual-dictionary cooperative occlusion error model was established.First,for the low-rank decomposition algorithm,the singular value of the kernel norm fluctuates greatly between different dimensions,which cannot effectively estimate the minimum rank of the matrix.The log determinant function is introduced to solve the problem of minimizing the rank of the face feature matrix.The estimation accuracy and the sensitivity of noise are included.The non-convex logarithmic low-rank algorithm is applied to the expression image to decompose the low-rank part and the sparse part,which removes the influence of individual differences on facial expression recognition.Then,the K-SVD algorithm is used to perform dictionary learning on the low-rank part and the sparse part,respectively,to obtain class-specific dictionaries and differential structure dictionaries.Finally,a separable subspace model,occlusion error,is proposed.It does not belong to any other subspace and does not contain any class-specific basis vectors.The dual dictionary collaborative error matrix representation classification method makes the spatial expression adaptive and expressive.The ability is stronger,and at the same time,the connection between each type of expression sample is described in more detail.(2)In view of the imbalance of the training data of the expression data set in the nonconstrained environment,and the large difference between facial expression images of the same kind,an expression recognition framework combining feature learning and semantic attributes is proposed.First,deep neural networks are used to learn the visual features of facial expression images.Then,based on the AU unit hierarchical analysis model,the semantic attributes of the expression image are established.Finally,the mapping relationship between visual features and semantic space is used for classification and recognition.The facial expression recognition method based on semantic attributes pays more attention to the advanced features of facial expressions,uses more distinctive and conceptual features to describe samples,and reduces the differences between samples in the class.In addition,constructing information sharing semantic attribute vectors embedded in visual features can expand the differences between classes and reduce the individual differences within the classes.