Research On Micro-expression Recognition Based On Method Of Local Crucial Region And Feature Selection

Micro-expressions are short expressions that appear on the face when people try to suppress or hide real emotions in a particular situation.It is the conscious action of people's psychological activities,which can often reflect the real emotional changes of human beings.They have important applications in many fields.However,microexpressions have the characteristics of short duration,small intensity and local occurrence,which makes the research of micro-expression using computer automatic face great challenge.According to the characteristics of micro-expressions,the main work of this thesis is summarized as follows:(1)There are also some unrelated muscle movements in the face while producing micro-expressions.The existing global method of micro-expression recognition extracts features of these unrelated actions,which will influence the recognition effect.According to the region of the action unit involved in the micro-expression,a research method based on local region is proposed,which divides the seven local regions related to the microexpression by the key point coordinates of the face.Firstly,the performance of microexpression recognition based on local facial features is studied.and it will reveal that there is a highly correlation between the eye area and surprise or disgust,and the similar correlations also lie in the mouth area and the happiness,the chin area and the depression.Then,the feature vector of local area combinations are extracted to perform microexpression recognition.The results show that the effect of micro-expression recognition in joint local region is better than that of global region method.(2)Feature descriptors LBP-TOP,HOG-TOP and HIGO-TOP are commonly used feature extraction methods for micro-expression recognition.However,such feature vectors have high dimension,large computation complexity,long running time and low recognition accuracy.The feature selection including information gain and Fisher feature selection are applied to reduce the dimension of feature vectors and improve recognition rate.The “Leave-One-Subject-Out Cross Validation” method is used to conduct the micro-expression classification experiment using the three kernels(Linear Kernel,Chisquare Kernel and Histogram Intersection Kernel)of support vector machine.A large number of experimental simulations show that two such feature selection methods can reduce the computational complexity,shorten the running time and greatly improve the recognition effect of micro-expression.