Dual Temporal Scale Convolutional Neural Network For Micro-Expression Recognition

Facial expressions play an important role in our daily-life and natural communications.Typically,facial expression is referred as macro-expression,which lasts from 1/2 seconds to 4 seconds and is easily recognized and identified.However,macro-expression can be misleading,it can be employed to cover up genuine emotions.Micro-expression,as a spontaneous facial movement and can leak the concealed emotion,has received extensive attention in recent years.Micro-expression is an uncontrollable,short facial expression that reflects the emotions of people trying to hide and the unconscious feelings that people do not realize,so it is more real and reliable to recognize human emotions through micro-expressions.The recognition of micro-expression is very challenging,because it is born with short duration period(last less than 1/2 seconds)and low intensity,while people can hardly perceive micro-expression,and it is also difficult to classify various micro-expression clips using pattern recognition methods.Additionally,due to the difficulty in collecting and labeling micro-expression data,the database that we can use so far is very limited.It is difficult to train an effective micro-expression algorithm based on effective database.To solve the aforementioned problems,this thesis proposed a m icro-expression recognition approach by using Dual Timing Scale Convolutional Neural Network(DTSCNN).This method first extended the micro-expression data set(CASMEI、CASMEII)to reduce the risk of over-fitting in the network training process,and then used the DTSCNN to extract the micro-expression video sequences at 64 fps and 128 fps time scale,and finally used SVM to conduct decision-level fusion and classification.DTSCNN not only solves the problem that the micro-expression database is small and difficult to train,but also achieves very good results: the recognition rate of DTSCNN(66.67%)on the CASMEI and CASMEII databases is higher than the traditional micro-expression recognition algorithm(MDMO: 55.45%,FDM: 56.97%,STCLQP: 56.36%).The experimental results show that DTSCNN is effective in recognizing micro-expression and achieves 10% improvement in the recognition rate comparing to the state-of-the-art traditional algorithms for recognizing micro-expression.