The Research Of Micro-Expression Recognition Based On Few-shot Learning

Micro-expression is a facial expression spontaneously inadvertently,which can better express people's real emotional changes than macro-expression,they are a very accurate and effective behavioral clue.Therefore,Micro-Expression have a wide range of application value in many fields,such as in education,interrogation,polygraph,anti-terrorism,clinical diagnosis,etc.However,due to the short duration of the micro-expression and the small motion range,it is difficult to be captured and recognized by the human eye.Therefore,it is of great significance to realize automatic and efficient micro-expression recognition.Early micro-expression recognition mainly relied on traditional image recognition methods requires a complicated manual feature extraction,and the recognition accuracy is low,so a micro-expression recognition method based on deep learning has emerged.Previous deep learning methods mostly used to extract spatial information to identify micro-expressions,or first to extract spatial information and then to extract time information,which split the connection between the two to a certain extent.At the same time,deep learning technology requires a large amount of data to train the neural network,but due to the influence of the characteristics of micro expressions,the data set and quantity of micro expressions are limited,which leads to poor recognition effect.In order to solve the problems of low precision in micro expression recognition,this thesis has completed the following research work.(1)Propose a micro-expression recognition model based on three-dimensional convo lutional neural network(Micro Exp3DCNN).Considering the time characteristics of microexpressions,this thesis uses 3D convolutional neural network to recognize microexpressions by extracting the spatial-temporal characteristics.The experiment was carried out on the two data sets of CASME? and SMIC-HS respectively,the influence of the size and number of different convolution kernels on the recognition effect was analyzed at the same time,and compared with the results of the previous micro-expression recognition method,it proved the effectiveness of the method proposed in this paper.And finally the recognition effect are visually presented in the form of a confusion matrix.The experimental results show that the proposed method can obtain better recognition results by extracting spatio-temporal features than the state of art,but due to the uneven classification of the data set samples,the recognition accuracy rate for classification with a small sample size needs to be improved.(2)In order to solve the above-mentioned problem of low accuracy of recognition with a small sample size,this thesis proposed to use the meta-learning algorithm on the threedimensional convo-lutional neural network(3DCNN)for small-sample learning for microexpression recognition.The prior knowledge is used to adjust and optimize network parameters and learn a good initial weight,so as to solve the problem of small sample learning of micro-expressions.In the training part of the micro-expression recognition network based on meta-learning,first the division of the meta-training set and the verification set is explained,and then the training process of the model is described,and the calculation of the meta-task loss value is explained to obtain the optimal parameters during training.The experimental data shows that the effect micro-expression recognition of 3 DCNN based on meta-learning algorithm has better accuracy and robustness than Micro Exp3 DCNN,and can achieve better recognition results for classification with a small sample size.