Micro-expression Recognition Based On Dynamic Sequence

Facial expression recognition is an important field. It is already very mature afternearly half a century of research. In recent years, people begin to pay close attention to akind of very special expression, which is called micro-expression. Due to the universalityand importance of its application in various fields, micro-expression becomes a hot directionof study.The work of this paper aims at the related issues of dynamic sequence micro-expression recognition. The main contents are as follows,1. The micro-expression sequence image preprocessing. The research of this paper isbased on the SMIC database. Relevant image processing is carried out in view of the natureof the database. First, the gray scale normalization and scale normalization are introduced.Second, since the length of the sequence in the database differs from11to58, timeinterpolation algorithm was used here to normalize the sequence to the same length.Experiment results showed that time interpolation algorithm can change the length of thesequence to any number under the precondition of undistorted. It provides a stable conditionfor subsequent processing.2. Micro-expression feature extraction. Sequence feature extraction is mainly to obtainthe relationship between images and the corresponding texture changes. This part analyzedseveral typical dynamic sequence feature extraction algorithms. LBP-TOP method was usedas the feature extraction method according to its advantage in texture extraction. First, thelocal binary pattern of static image was extracted. The results show that, local binary patternhighlights the texture characteristics of the major organs. Second, local binary pattern wasextended to three orthogonal planes and the local binary pattern of sequence was got.3. Magnify the intensity of weak expression. There are low intensity expressionsequences in the database. Useful information in small movement is few. It may not be ableto highlight the characteristics of the expression type which is affected by subjects’ personalcharacteristics. In order to solve this problem, feature point motion amplification algorithmis applied to the sequences of weak intensity. Different amplifier parameters are chosenaccording to the characteristics of the various types of sequences. These samples aresynthesized to corresponding exaggerated sequences.4. Implementation of expression classification. Expression classification is the last stepof facial expression recognition, and it is a very important step. Due to the diversity andcomplexity of expression presentation, it is difficult to meet the requirements of the ideal classification with the linear classifier. Through the analysis of several classificationmethods, combining the features extracted in this paper, the SVM as classification algorithmwas selected as the recognition method in this paper. First, suitable SVM parameters arechosen through some experiments. Then, a large number of experiments are carried out toverify the performance of the method used in this paper. Comparative experiments provedthat the algorithms used in this paper have achieved satisfactory results.