Research On Micro-expression Recognition Algorithm Based On Deep Forest

Micro-expressions contain a wealth of emotional information,which is produced by humans trying to suppress their own real emotions.It has broad application prospects in criminal investigation,psychology,clinical medicine,etc.Recognizing microexpressions quickly and accurately can predict a person’s true mental state to provide necessary technical support for different fields.This thesis focuses on the use of deep forests for micro-expression recognition research,and solves the problems of low microexpression recognition rate and poor interpretability of deep models.The main work is as follows:(1)This thesis studies the interpretability of the deep forest,embeds the autoencoding forest into the deep forest,makes it use the decision path of each subtree to visualize the features captured by each neuron,and make the model interpretable,restore the classification process,solved the problem of poor interpretability of deep models.(2)This thesis proposes an interpretable deep forest micro-expression recognition algorithm based on error correction output codes.On the basis of deep forest,one-to-one and one-to-many ECOC units are used to enhance the algorithm’s fault tolerance.The auto-encoding forest is used as the base classifier of ECOC.Different class decomposition schemes in two units offer us a deep insight into the difference among ME classes,so that they can be better applied to the task of micro-expression recognition.(3)A deep forest-based micro-expression recognition based on dynamic channels and reverse iteration is designed.First of all,to address the problem of excessive redundant information of micro-expression features,important feature inputs are selected through random forest and complete random forest,and the model divided into four channels,input the vertex frame,horizontal vector,vertical vector,and important features of light strain in the micro-expression video to train the model.Each layer uses the datarelated ECOC matrix,according to the class interval between data encoding.After the adaptive growth of the model is completed,the concept of dynamic channel is used to remove the poorly effective channels,and the ECOC code matrix is modified in the reverse iteration to train difficult categories,which improves the performance of micro expression recognition and reduces memory.