Research On Sparse Representation Algorithms For Facial Expression Recognition Based On Decoupling Of Feature Learning Space

With the rapid development of computer and artificial intelligence technology,the technology of facial expression recognition has a wide application prospect.Although many facial expression recognition algorithms have achieved good performance,facial expression recognition system still faces many challenges in practical application.The incomplete training data is usually a fusion image of face identity and expression feature,which will affect each other.Therefore,how to decouple effective expression features and suppress the impact of face identities on the expression features,and designing an automatic expression recognition system that is independent to face identities has important theoretical and application value.This paper starts from the aspect of feature extraction,and constructs recognition models based on the decoupling space feature learning based on the characteristics of the incomplete data fuse face identity and expression feature.Feature extraction approaches are based on the shallow features and the deep features,which are based on the prior knowledge and the self-learning decoupling spaces,respectively.The specific research contents are as follows:(1)According to the incomplete training data simultaneously fuse face identity and expression features,individual-free representation based classification for facial expression recognition and dictionary learning feature space via sparse representation classification for facial expression recognition are proposed.Individual-free representation based classification for facial expression recognition first transforms the original space into expression space and extends the incomplete training samples to make up the influence caused by small sample size.The algorithm uses collaborative representation for the final classification.Similarly,dictionary learning feature space via sparse representation classification for facial expression recognition first transforms the original space into expression space.Then this algorithm uses the symmetry structure of image to learn the discriminative dictionary and uses the principal component analysis to reduce dimensions.Sparse representation is used for the final classification.(2)According to lacking of a neutral expression set for training samples that cannot be used to construct an expression dictionary with this neutral set,robust facial expression recognition with low-rank sparse error dictionary based probabilistic collaborative representation classification is proposed.The original training dictionary is first converted into a set of representative bases with a corresponding LR common dictionary and a low-rank sparse error LRSE dictionary.Then the test sample can be represented and classified by probabilistic collaborative representation classification.(3)Discriminative feature learning based pixel difference representation for facial expression recognition is proposed from the perspective of data.This algorithm tries to improve the representation ability of expression features by using a discriminative feature descriptor.We obtain the discriminative feature dictionary based on pixel difference representation.Next,we use vertical two-dimensional linear discriminant analysis in direction for reducing dimensions.Finally,we use the nearest neighbor classifier to determine the labels of the query samples.(4)In view of the excellent feature extraction ability of deep subspace model,combing the kernel collaboration representation and deep subspace learning for facial expression recognition and a two-phase representation classifier driven deep subspace learning for facial expression recognition algorithms are proposed.The former algorithm tries to map these features to kernel space to effectively their non-linear similarities.The collaborative representation method finally is used for classifying.Tthe latter algorithm uses the Euclidean distance to select the training samples that is close to the test sample for the representation and classification from the perspective of feature selection and improves the recognition performance.