Research On Micro Expression Detection And Recognition Method Based On Optical Flow Features

Micro-expression is a kind of expression that people can't help but quickly present when they try to control and hide their strong emotions.Because of the uncontrollability of micro-expression,we can identify the real emotion hidden by people through the observation of micro-expression.Micro-expression research has great potential in clinical treatment,crime investigation,business communication and other fields.Because of the short duration and small amplitude of micro-expression,it is very difficult for human to detect and recognize micro-expression.Compared with human beings,computer has unique advantages in detecting and recognizing micro-expression.High frame number camera can effectively capture and record the whole change process of micro-expression,and migration model can help to promote mature micro-expression detection and recognition model in a wide range.Therefore,in recent years,the demand for automatic facial micro-expression recognition is increasing.However,there are still some problems in the research of micro-expression:(1)the existing micro-expression detection algorithms are vulnerable to noise,and the false positive rate is high.In the process of micro-expression happening,head translation will have a great impact on micro expression feature extraction.(2)In the actual application process,due to the different acquisition environment and equipment,the distribution of micro expression samples in training set and test set is different,which will affect the results of micro expression recognition.And the sample of single micro expression data set is small,only using a single micro expression data set as the training set may make the model training insufficient(3)There is a lack of systematic analysis of micro expression movement in the existing research.The main research contents of this paper are as follows:(1)aiming at problem1,the improved optical flow method is used to accurately locate the facial movement and effectively improve the accuracy of micro-expression features.The optical flow features of 14 regions of interest are extracted and the feature curves are drawn.The start frame,peak frame and end frame of micro-expression are located from the waveform change of characteristic curve.(2)To solve the second problem,two multi-source domain adaptive methods are used to reduce the distribution difference between domains,and solve the influence of different data acquisition environment and equipment in practical application.(3)Aiming at the third problem,the self-organizing mapping network(SOM)is used to cluster the optical flow characteristics of micro-expressions,and the motion rules of different kinds of micro-expressions are analyzed and summarized.On this basis,the visualization method of cluster center parameters is studied,so as to intuitively and clearly present the overall motion trend and change rules of various micro-expressions.In order to meet the needs of the practical application of micro expression research,I combine the three parts of the work,namely,micro expression detection,micro expression recognition and micro expression movement trend visualization.A simple graphical interface is designed to show the three steps in the form of buttons and images,which meets the needs of real life.In this paper,the proposed method is evaluated on CAS(ME)~2and SAMM Long Videos datasets,using accuracy,recall and F1-score as evaluation criteria.The F1-score of CAS(ME)~2and SAMM Long Videos datasets are 0.3530 and 0.3690,which are improved by 0.2127 and 0.0391,respectively.In this paper,CASMEII,SAMM,SMIC datasets are used as the evaluation datasets of micro-expression recognition algorithm.In order to implement a unified evaluation index,the emotion classification labels in all three datasets are mapped to negative,positive and surprise.Unweighted F1 parameter(UF1)and unweighted accuracy(UAR)were used as evaluation criteria.UF1 and UAR are 0.6669 and0.7089 under the condition of unsupervised single source domain;UF1 and UAR are 0.7213 and 0.7786 under the condition of unsupervised muliti-source domain;UF1 and UAR are 0.7707 and 0.8039 under the condition of supervised multi-source domain.