Research On Micro-expression Recognition Assisted By Macro-information

As the true representation of one's inner emotion,micro-expression is an instantaneous real reaction that people involuntarily show when they are strimulated.Micro-expression has the characteristics of short duration,low intensity and difficult to induce.Since the occurrence of micro-expression is not controlled by the mind,micro-expression can reveal the real emotion,even if a person tries to hide the real emotion,the movement and contraction of facial muscles can reveal the true emotions.In the case of lie detection,micro-expression recognition can be used to determine whether the speaker lies or not during the conversation,which has great application value and development potential in national security,judicial trial,airport security,criminal detection and prison management.Micro-expression can help to maintain the stability and security of the country.Automatic micro-expression recognition technology can automatically recognize micro-expressions to estimate people's inner real emotions,which has broad prospects of development.With the development of machine learning and pattern recognition,micro-expression automatic recognition technology develops rapidly.However,there are too few marked training samples from the existing micro-expression database,thus a high precision model cannot be trained,leading to an unsatisfied performance of micro-expression automatic recognition.Current macro-expression and speech databases have enough labelled samples,simultaneously,macro-information(macro-expression/speech)and micro-expression have certain emotional characteristic similarities.Therefore,in order to solve the problem of recognition difficulty caused by the limited marked sample of micro-expression,in this dissertation,we propose micro-expression recognition algorithms assisted by macro-information.In all,the main contribution of this dissertation is listed as follows:The related concepts of micro-expression and micro-expression recognition are expounded.Then the background and significance of the research are summarized.Besides,the research status in the field of micro-expression recognition at home and abroad is reviewed.A micro-expression recognition model based on singular value decomposition(SVD)is proposed.Macro-information is combined with micro-expression by means of SVD,which achieves the transfer learning from macro-information to micro-expression.This model makes full use of a sufficient number of macro-information samples,solving the problem of low recognition accuracy due to the lack of micro-expression labeled training samples.In addition,the effectiveness of the model is verified by using CK+ expression database,CASMEII micro-expression database and CASIA Chinese emotional corpus.A micro-expression recognition model based on coupled metric learning is proposed.New macro-expression and micro-expression feature descriptors are proposed to enhance the accuracy of feature description,which include Dual-cross Patterns from three orthogonal planes,Multiple order Dual-cross Patterns,Multiple order Dual-cross Patterns from three orthogonal planes,Hot Wheel Patterns and Hot Wheel Patterns from three orthogonal planes.Due to the closer distance between macro-information and micro-expression in a common space,coupled metric learning algorithm is deployed to project macro-information and micro-expression into the common subspace.Then Smooth Support Vector Machine is applied to proceed classification.The experimental results show that the proposed micro-expression recognition model based on coupled metric learning has better performance than other micro-expression recognition models.A micro-expression recognition model based on macro-information knowledge transfer is proposed.In order to enhance the nonlinear representation of macro-information and micro-expression features,a kernel coupled discriminated local block alignment algorithm is proposed,which can project macro-information and micro-expression into a common space after kernel mapping.For classification,a transfer support vector machine is proposed,which takes advantage of the structural risk minimization criterion of statistical theory to adjust support vector by the method of weighting.Micro-expression recognition rate can reach the top of 91.5%by the micro-expression recognition model based on macro-information knowledge transfer.