Research On Facial Expression Recognition Based On Selective Ensemble Optimized By Bee Colony

Facial expression recognition involves many subjects,such as psychology,computer vision,pattern recognition,and so on.It has high theoretical value and practical significance.How to use machine learning and deep learning to improve accuracy and timeliness is a research focus in recent years.In this thesis,a selective ensemble expression recognition method based on artificial bee colony algorithm is proposed,and an expression recognition method based on ResNet and selective ensemble is proposed.In pattern classification,compare with single classifier and ensemble learning,selective ensemble can achieve better result,and different optimization algorithms also makes different effect.In this thesis,a selective ensemble expression recognition method based on artificial bee colony algorithm is proposed.In the pattern classification,adopt artificial bee colony algorithm to allocate different weights to the ensemble classifiers,and eliminate the classifier with little weight.Comparison experiments with single classifier,integrated learning and different optimization algorithms on JAFFE,CK+ and KDEF datasets show that the proposed method achieves better results of accuracy and global optimization.Then,on the basis of the selective ensembe proposed above,a "ResNet and selective ensemble" expression recognition method is proposed.ResNet is a new convolution neural network,It has strong ability of feature extraction and generalization.This thesis analyzes the structure and advantages of the ResNet.Then,we use ResNet to extract the feature,and compared the effects of 18 layers of ResNet,34 layers ResNet and traditional feature extraction methods.The advantages of the ResNet are proved.In addition,this thesis analyzes the difference between softmax in ResNet and SVM classifier.In order to make full use of the advantages of selective ensemble SVM,the features extracted from the ResNet are classified by selective ensemble SVM,and then,compare with the softmax.Comparison experiments on JAFFE,CK+ and KDEF datasets show that the proposed method achieves higher recognition rate.