Research On Facial Expression Image Recognition Based On Feature Fusion And Voting Model

The field of human-computer interaction is a research hotspot in social science,and the continuous progress of computer technology has spawned more and more new formats of human-computer interaction.In recent years,with the rise of smart device wearing and cell phone face unlocking,facial expression recognition has become a research hotspot of human-computer interaction.This paper mainly studies the extraction of facial expression fusion features,improves two excellent convolutional neural networks,and combines the advantages of machine learning and deep learning.Then analyze and compare through experiments.The main work of this paper is as follows:1.Cut and normalize the facial expression database.In this paper,CK+,JAFFE and FER2013 databases are used,Adaboost algorithm is used to roughly cut out the facial expression images,and scale normalization method is adopted for the cut-out facial region images to obtain the facial expression images with the same size.Finally,compare the advantages and disadvantages of multi-dimensional classification expression feature extraction methods and classification algorithms.2.This paper proposes an expression recognition method based on Dlib feature fusion.Two classical algorithms LBP(Local Binary Pattern)and Gabor Wavelet Transform are selected for feature extraction of facial expressions considering learning rules and facial expression features.The optimal classifier is selected from SVM and softmax,and the two algorithms are cross-tested with the two classifiers.The Dlib database used for face recognition in Python is introduced to extract 68 feature points of face.The features extracted by LBP and Gabor wavelet transform are uniformly sampled and dimensionalized,and then fused with the 68 feature points extracted by Dlib,which are put into the optimal classifier SVM for classification.Finally,a feature fusion model suitable for facial expression recognition and classification determined by experiments is obtained.In CK+ database,the highest correct recognition rate of this model is 96.0%,and in JAFFE database,the highest correct classification rate is From the accuracy rate,we can see the superiority of this model for facial expression recognition.3.A voting model based on improved convolutional neural network is proposed.Two convolution neural network algorithms VGG19 and Resnet18,which perform well in facial expression experiments,are improved by changing the pooling mode to pyramid pooling and changing the convolution kernel size from 3×3 to 1×1,which improves the model accuracy.The voting model is composed of VGG19,Resnet18,LBP and Dlib feature fusion models,which verifies the accuracy and effectiveness of the voting model in facial expression recognition.When faced with two small databases JAFFE and CK+,the accuracy rate is 98.59% and 99.64% respectively,and when faced with the large database FER2013,the accuracy rate is 74.58%,which has obvious advantages over other recognition methods.The voting model solves the problem of universality for databases of different sizes,and achieves excellent classification results.