Research And Implementation Of Multi-view Facial Expression Recognition Based On Generative Adversarial Networks

Multi-view facial expression recognition has always been one of the research difficulties in the field of machine learning and computer vision.Facial expression recognition in the traditional sense refers to the classification of positive facial expression pictures.However,in reality,most of the facial expression images we acquired are multi-view facial expression images taken by cameras from various angles.The characteristics of the same facial expressions at different angles are quite different,making the task of recognizing multi-view facial expressions by computers more challenging.In addition,many public facial expression recognition algorithms have poor effect on face image recognition under non-restrictive conditions,indicating that most of the current expression recognition algorithms still have a large gap from practical applications.Most of the expression recognition algorithms only classify the standard face expression images by related algorithms.If the traditional face expression recognition algorithm is directly used in the multi-view face expression recognition field,the accuracy rate will be greatly reduced because of the difference between the features of frontal face expression and multi-view facial expression.In view of the above problems,this paper adopts an algorithm based on GAN to improve the accuracy and robustness of multi-view facial expression recognition.For the facial expression pictures collected at different angles,based on the idea of transfer learning,the multi-view facial expression features are mapped into the positive facial expression feature space,and the method of feature classification is optimized.The algorithm can well learn the mapping relationship between multi-view facial expression features and frontal facial expression features,and can extract the spatial features of facial expressions from two angles.The effectiveness of the algorithm is fully verified in the experimental part of Chapter 4.The research content of this paper is as follows:(1)Investigated the commonly used multi-view facial expression recognition method,briefly described the relationship between multi-view facial expression recognition and migration learning,and used the deep neural network-based algorithm to realize three experiments in the field of transfer learning.It proves the feasibility of generating anti-networks in the field of migration learning;(2)An algorithm based on CGAN is proposed to solve the multi-view facial expression recognition problem.In recent years,the generation of GAN has achieved great success in the field of computer vision.This paper introduces it into the field of expression recognition and has achieved good results.In order to improve the accuracy of multi-view facial expression recognition,this paper uses the branch of the GAN--CGAN to learn the mapping relationship between the frontal facial expression features and the multi-view facial expression features.Converted multi-view facial expression to the corresponding frontal face expression,and then use the CNN to classify the expression of the obtained frontal face expression.The algorithm uses the idea of transfer learning to learn the mapping relationship between the frontal facial expression features and the multi-view facial expression features,which is correct the camera angle,and then simply recognizes the frontal facial expression.This paper discusses the use of CGAN to combat the network for multi-view expression recognition system,mainly in the form of experiments;(3)A deep neural network named MVFGAN combining the advantages of GAN and CNN is proposed.The lower layer portion of the feature mapping network GAN is connected to the high-level portion of the expression discrimination network VGG16 to form a deep neural network MVFGAN for discriminating the facial expression class.Since the lower part of the GAN is the mapping of the side face image to the face image,the upper part is a description of the expression semantics in the face expression image.Therefore,the method utilizes GAN to decompose the multi-view facial expression recognition problem into a side face image to a face image and a frontal face expression image.The network and the use CGAN for multi-view expression recognition are different,that is,it is not necessary to generate a frontal facial expression image corresponding to the multi-view facial expression image in the middle and then use the CNN to classify,but to classify the feature map and the feature.The two networks are coupled to form a new end-to-end deep neural network MVFGAN.After using the MVFGAN algorithm for multi-view facial expression recognition experiments,the two network models are merged in view of the complementarity of the algorithm and the CGAN-based algorithm,which further improves the accuracy of multi-angle facial expression recognition.Finally,the proposed MVFGAN fusion algorithm and the existing multi-view facial expression recognition algorithm are compared on several well-known facial expression datasets.The proposed method is superior to other methods,and effectiveness of the algorithm is also proved.