Font Size: a A A

Research And Implementation Of Multi-view Facial Expression Image Generation And Recognition

Posted on:2021-04-25Degree:MasterType:Thesis
Country:ChinaCandidate:H QinFull Text:PDF
GTID:2428330614965898Subject:Facial expression recognition
Abstract/Summary:PDF Full Text Request
Multi-view facial expression recognition has always been one of the difficulties in the field of image processing.Traditional facial expression recognition is mostly based on positive facial expression images.However,in real life,most of the facial expression images we acquire are taken by cameras with different angles.The facial expression characteristics at each angle are very different,making it difficult for the computer to recognize multi-view facial expressions.In addition,the current mainstream expression recognition algorithm has a high recognition rate for positive facial expression images,while the multi-view facial expression recognition image recognition effect is not good,indicating that multi-view facial expression recognition is a subject worth exploring.Multi-view facial expression recognition is faced with a series of problems such as fewer databases and higher cost of manual facial expression annotation.Regarding the issue above.In this paper,the multi-feature fusion loop generation adversarial network MFCGAN is used to improve the accuracy and robustness of multi-view facial expression recognition.The design idea of the network is to obtain a classifier that can accurately make multi-view side facial expression judgments through the labeled front face images without using the side facial expression category tags.Among them,the idea of transfer learning is used to map facial expression features from different angles to the facial expression feature space,and the expression recognition network is improved to improve the accuracy and robustness of multi-view side facial expression recognition,and has huge development prospects.The experimental part of Chapter 4 of this paper can fully verify the effectiveness of the algorithm.The research contents of this article are as follows:(1)The principle and learning algorithm of generative adversarial network are introduced in detail,and some derivative models of generative adversarial used for image analysis in recent years are introduced,which proves the feasibility of the application of generative adversarial network model in face image analysis(2)The MFCGAN network is proposed to solve the multi-view facial expression recognition problem.Extract the key points(eyes,nose,mouth)of the side faces and creatively combine the side face pictures and the extracted key information of the side faces into the forward face image generator in the form of a four-channel heat map In the process,the front face image is synthesized from the side face image,and then the generated front face picture is input into the reverse face generator to reconstruct the corresponding side face picture and compare it with the original side face picture until the discriminator cannot distinguish Whether it is a real side face picture or a generated side face picture,thus forming a multi-feature fusion cyclic generation adversarial network.This paper discusses the system of multi-angle expression recognition using multi-feature fusion cyclic generation adversarial network,mainly in the form of experiments;(3)A new hybrid network structure HFRNET is proposed to perform multi-view facial expression discrimination on side faces.The trained MFCGAN generator is used as the shallow part of the HFRNET network,and the deep part of the positive face expression classification network is used as HFRNET.The deep part of is constructed a deep neural network HFRNET used to judge the facial expression category.Therefore,the multi-view expression recognition problem is decomposed into the mapping of side-face images to front-face images and the discrimination of front-face expression images.The main advantage of this network is that there is no need to use the convolutional neural network to classify the positive facial expression pictures corresponding to the side facial expression pictures.Instead,the two networks of feature mapping and feature classification are coupled to form a new end-to-end Deep neural network HFRNET.After the experiment of multi-angle facial expression recognition using the HFRNET algorithm,in view of the complementarity of the algorithm and the MFCGAN network algorithm,the fusion of the two network models further improves the accuracy of multi-view facial expression recognition.This paper overcomes the limitation that traditional paired data sets must be used for training in the training process of traditional generative adversarial networks.The multi-feature fusion-based cyclic generative adversarial network is used to achieve multi-angle people when there are no facial expressions in the training.By comparing with existing multi-view facial expression recognition algorithms in RAFD and Multi-PIE data set,the method proposed in this paper is obviously superior to other methods,which can confirm the effectiveness of the algorithm in this paper.
Keywords/Search Tags:Multi-view facial expression recognition, cycle generation adversarial network, local facial expression feature interception
PDF Full Text Request
Related items