Research On Facial Expression Recognition Based On Deep Learning

With the coming of intelligent era,facial expression recognition technology has more and more practical value in many fields.The continuous development of deep learning brings new breakthrough and development trend for facial expression recognition.Facial expression recognition has always been a challenging task,which is seriously affected by low quality face image occlusion and feature redundancy.Based on the deep learning technology,this thesis innovatively proposes that the convolution neural network and the generation countermeasure network will be combined to improve the low-quality data;in the feature extraction stage,the clustering algorithm is used to screen the deep features,which greatly improves the accuracy of expression recognition.The details are as follows:1.To solve the problem of occlusion or breakage of low-quality face image,an end-to-end facial expression recognition method is proposed.In this thesis,the existing face data set is manually added with damage or occlusion as a low-quality sample set.The restored image,the low-quality facial expression image and the original image are combined into a three input classifier by convolution neural network classifier.The training of the model is guided by the three element sorting constraint to achieve better classification effect.2.To solve the problem of feature redundancy,a method of feature graph clustering based on convolution neural network is proposed.Firstly,the feature map of the last convolution layer of the network is extracted by the pre training network,and then the cluster operation is carried out.The cluster center is taken to form a new set of feature map,and the classifier is trained by the cluster feature map.In this thesis,the supervised deep learning method is combined with the traditional machine learning method,and the feature graph clustering is used to remove the redundancy of features so that the network can learn more effective features.For facial expression recognition,the method 1 proposed in this thesis solves the problem of input image data,and the problem 2 solves the problem of feature redundancy in the process of feature extraction.The methods in this thesis are verified on fer2013 and CK + facial expression data sets.The method based on the generation of antagonism network has a high recognition rate on low-quality facial expression,and the use of ternary sorting loss can better guide the network to repair the image;the depth neural network can achieve a relatively high classification accuracy through feature map clustering.