Research On Micro-Expression Recognition Technology From Construction To Deep Self-Learning

Micro-expression recognition has become a hot topic in the field of computer vision in recent years,and micro-expressions have important applications in lie detection.So far,the task of micro-expression recognition still faces the challenge of low recognition rate,insufficient data sets caused by the difficulty of collecting the data set,and a very unbalanced number of samples in the spontaneous micro-expression data sets caused by the difficulty of inducing various micro-expressions.This article transitions from using traditional feature extraction methods to using deep learning methods to extract features of micro-expressions.Specifically,the main contributions are as follows:Firstly,aiming at the problem that not all facial regions have a positive effect on micro expression recognition,the Local Binary Patterns from Three Orthogonal Planes(LBP-TOP)of traditional feature descriptors are improved,a micro-expression recognition algorithm based on multi-scale fusion block-based LBP-TOP is proposed.First of all,the algorithm follows the LBP-TOP baseline method to divide the facial regions according to the fixed grid,then a single facial region is used for classification,and the average value of the classification accuracy of all facial regions is taken as the threshold to select effective blocks to selectively merge local features.Considering that some global information may be lost,select more facial sub-blocks to extract features and select effective blocks,multi-scale fusion of global features.By selecting effective blocks,the feature dimension is reduced,and the efficiency is improved faster and the recognition accuracy is improved.Secondly,the Learning-based Motion Magnification(LBMM)network is used to magnify the tiny movements in the micro-expression frames sequence,and then the peak frame is selected from the magnified sequence as the input of the improved Residual Neural Network(Res Net).The improved Res Net18 network divides the output features of the last convolution layer of the Res Net18 network into upper and lower parts by fusing part-based mechanisms.The two parts and their fusion features jointly learn,and use multi-objective functions to optimize the network.Solved the problem of low recognition rate due to low micro-expression intensity.Finally,a total variation network(TVnet)is introduced to train the optical flow.During the training,a fixed onset frame(reference frame)is required as one input of the TVnet network,and the other input selects the frames in the micro-expression sequence in turn,by large sample training TVnet network to ensure accurate optical flow information.Then from TVnet network output optical flow sequence,select the optical flow information obtained from the peak frame and the onset frame as the input of the improved Res Net18 network.Based on the part-based mechanism,attention network are fused to select more discriminative features.It solves the problem that micro-expression is difficult to recognize due to its short duration.Experimental results show that the methods in this paper have achieved better recognition accuracy than the previous methods.