Research On Cross Source Image Recognition Technology For Micro-expression Recognition

Emotions are people's attitudes towards events or things around them,and facial expressions are one of the most powerful,direct and natural ways for people to express their inner feelings and intentions.Facial expression mainly refers to the stress response of people after being stimulated,which includes macro-expression and micro-expression.The differences between macro-expression and micro-expression mainly lies in the duration and intensity.Compared with macro-expression,micro-expression is characterized by short duration,low intensity and difficulty to induce.Micro-expression is a kind of non-spontaneous and immediate facial dynamic during the emotional experience and can usually reveal the real emotions that people want to hide.Micro-expression detection and recognition have many important applications,such as crime detection,business negotiation,etc.,which have great application value and broad prospects in various fields.With the continuous development of computer technology and pattern recognition technology,facial expression recognition has attracted extensive attention,which has become an important research field of the new branch of science,and micro-expression recognition is an important research area in facial expression recognition.Due to the success of machine learning in the fields of classification,clustering and regression.micro-expression recognition is mainly realized by machine learning at the present stage.However,due to the limitation of the training sample size,the model training is not optimized enough and the recognition rate is not high.In traditional machine learning,both training and test data need to come from the same feature space,while cross-source learning can combine all tasks and then learn systematically,which solves the problem of different data distribution and,to some extent,can solve the problem that model training is not optimized due to insufficient data.Due to the correlation between macro-expression and micro-expression in emotional characteristics,a cross-source micro-expression recognition algorithm based on macro-expression is proposed in this dissertation to solve the problem that the model training of micro-expression recognition is not optimized due to the limited training sample size.In all,the main contributions of this dissertation are as follows:· A coupled source domain targetized with updating tag vector is proposed,which projects macro-expression and micro-expression iinto a common subspace,and make full use of information in source domain and target domain to realize cross-source learning from macro-expression to micro-expression.· A joint sparse dictionary learning decomposition for micro-expression is proposed.The macro-expression and micro-expression are projected into the tag space for dictionary learning,and their respective dictionaries are two-decomposed to find the common expression information part of macro-expression and micro-expression.The emotional connection between macro-expression samples and micro-expression samples is constructed to improve the micro-expression recognition rate.Experimental analysis shows that the proposed micro-expression recognition model based on joint sparse dictionary learning decomposition can effectively improve the micro-expression recognition rate.· A specific class of strong correlation tensor coupled metric learning model for multi-task multi-view learning is proposed,which takes advantage of the abundant expression samples,and learns the mapping of specific classes to map the macro-expression and micro-expression samples into the public space at the same time.Besides,trace norm is used to enhance the correlation between samples of specific classes.In order to reduce the redundant information introduced by multiple features,F norm and L21 norm are used for feature selection of multiple features.In addition,the micro-expressions are segmented around the feature points for multi-task learning.