Research On Micro-expression Database Establishment,Detection And Recognition

Different from the conventional facial expressions,micro-expressions are unconscious,transient facial expressions that reveal the genuine emotions people try to hide.Micro-expressions can provide information in many applications,such as lie detection and criminal detection.However,there are some difficulties in the study of it.In terms of micro-expression databases,the collection of micro-expressions is difficult.At the same time,there is no unified standard for database establishment.Therefore,there are few databases for micro-expression research,which greatly limits the development of the research.In the aspect of micro-expression detection,traditional detection techniques are only based on texture features or optical flow features,and the results are not ideal.In the aspect of micro-expression recognition,due to the lack of micro-expression databases and few samples of training set,both traditional and deep learning methods are limited to some extent.How to reduce the work of manual coding and obtain rich information that can be used for micro-expression recognition are the topics that need further research.In order to alleviate the lack of the existing micro-expression databases,we establish the SDU_spotting database dedicated to micro-expression detection.To address the problem of high labor and time consumption of manual coding,we propose a multi-domain fusion micro-expression detection model based on action units and an unsupervised micro-expression recognition algorithm based on distributed adaptation.Experiments are conducted on multiple micro-expression databases.Specifically,the contributions we made in this paper are as follows:Firstly,SDU_spotting database,which is specially used for micro-expression detection,has been established.Among the existing micro-expression detection databases,the SDU_spotting database has the largest number of samples and the highest resolution.In addition,SDU_spotting has been scientifically coded.It provides all samples' AUs,emotion types,and information about the start,apex and end frame.Secondly,a multi-domain fusion micro-expression detection model based on action units is proposed.In this model,the facial action unit sub-block that contains the most micro-expression action unit information are found.Then,the optical flow feature and frequency feature are combined to realize the fusion of time-space domain and frequency domain.Experiments on SDU_spotting and CASME ? databases show the effectiveness of this algorithm.Thirdly,an unsupervised micro-expression recognition algorithm based on distributed adaptation is proposed.Different from the general micro-expression recognition model,this algorithm introduces the informative macro-expressions knowledge.It uses the macro-expression and micro-expression databases to identify another micro-expression database without any labels.The algorithm consists of two parts:a source domain selection model and an adaptive distribution alignment model.A large number of experiments fully verify the effectiveness.