Research On Micro Expression Recognition Method Based On Interpretable Model

In recent years,deep learning technology has been developed and gradually applied to various industries with the improvement of computer hardware performance,facial recognition research based on deep learning has become a research hotspot.At present,the research of facial macro-expression has achieved excellent results,but the research of micro-expression recognition is still in its infancy.Facial micro-expressions have the characteristics of short duration and low intensity.Compared with facial macro-expression recognition,the difficulty of micro-expression recognition has been significantly increased.If using human to recognize micro-expressions,professional training of personnel is required,which is not only time-consuming and labor-consuming,but also difficult to ensure the accuracy.Therefore,how to make the computer automatically recognize micro-expressions is a problem worthy of study.First of all,this paper studies the Temporal Convolution Network based on 3D human action recognition.Since the input of the Temporal Convolution Network is one-dimensional skeleton feature sequence,and the micro expression video frames need to be input when performing the micro expression recognition task in this paper.And the network not only needs to have the ability to process time information,but also needs to be able to learn spatial information.Therefore,this paper proposes an interpretable spatiotemporal convolutional network model for micro-expression recognition.Based on the Temporal Convolution Network,spatial information is added to it,and the structure of the model is adjusted to make the improved model better applied to the task of micro-expression recognition.In addition,the activation maximization method is used to understand the features captured by the neurons in the hidden layers of the model,which makes the model interpretable.Secondly,this paper studies and analyzes the facial action coding system,and it is found that when micro-expressions occur,in most cases,only parts of the face change,and the magnitude of the change is obvious.Based on the phenomenon,this paper studies the application of the Attention mechanism in computer vision,and finds that the Attention mechanism can select a part of useful information from a large number of input information to focus on processing.Therefore,this paper combines the Attention mechanism with the original Res-STCN model,and proposes an interpretable micro-expression recognition method which combined with Attention mechanism.By adding the pixel attention module after the convolution layer of the residual unit of the Res-STCN model,which makes the model pay more attention to some key areas,improves the performance of the model in the task of micro expression recognition,and explains it by visualizing the weights of the pixel attention module to ensure the interpretability of the model.Finally,the proposed method is tested on the CASME2,SMIC and SAMM data sets,and compared with the existing research results.Experimental results show that the method in this paper is better than most current micro-expression recognition methods.