Research On Face Image Repair And Expression Recognition Method Based On GAN

Facial expressions contain rich personal emotional information,and automatic recognition of facial expressions has broad application prospects in the fields of human-computer interaction,intelligent security,and psychological analysis.At present,researchers at home and abroad mostly take the frontal unobstructed face image as the research object of the recognition algorithm.However,the face images taken in real life often have occlusion caused by glasses,hats and light,etc.,and the expression recognition of the blocked face image mainly uses the remaining features of the unoccluded area of the image for expression recognition,which greatly reduces The recognition rate and robustness of expression recognition algorithms are discussed.Therefore,how to design an algorithm that can solve the facial recognition of partially occluded facial images is a very research-worthy subject.In view of the above problems,this thesis proposes a GAN-based occlusion facial image repair and expression recognition algorithm,which decomposes the facial occlusion facial expression recognition problem into two stages,that is,first repairs the occlusion facial image,and then It performs facial expression recognition to accurately recognize the facial expressions of blocked faces.The research work of this dissertation is summarized as follows:1.This thesis draws on the idea of generating an adversarial network,and designs a face repair algorithm.The network structure of the algorithm is based on a dual discriminator to generate an adversarial network.The generator is based on the encoding-decoding symmetric structure of the VGG19 network.The discriminator consists of a local discriminator.It is composed of a global discriminator.The global discriminator is responsible for measuring the smoothness and authenticity of the overall structure of the image,while the local discriminator is mainly responsible for measuring the authenticity of image restoration in the occlusion area.Secondly,the adversarial loss function no longer uses the original adversarial function,but introduces an improved adversarial loss function based on the Wasserstein distance.The improved loss function is used to solve the disadvantages of generating adversarial network training is too free and difficult to control training.The ultimate goal of these improvements in this dissertation is to make the restoration effect of the restoration algorithm in this dissertation better,more realistic and smoother.Finally,through experiments,the repair algorithm in this dissertation is evaluated in two aspects: subjective vision and objective evaluation system.The experimental results show that the repair effect of the repair algorithm in this paper is remarkable,and the repaired face image is very realistic and close to the real image.For the image with a small occlusion area,it can be almost completely restored,thus laying a solid part for the later facial recognition part the basics.2.Combined with GAN-based facial image repair algorithm and convolution block-based expression recognition algorithm to design a partial occlusion facial expression recognition algorithm.The basic principle of the algorithm is to repair the occlusion face image first,and then recognize it.The repair part is replaced by a training network based on the GAN-based image repair algorithm.The recognition part uses a convolution block-based expression recognition algorithm to classify the expression features extracted by the five convolution blocks through the Softmax layer to achieve expression recognition.In the training stage of the algorithm model,it is pre-trained with a standard data set,and then the recovered face data set fine-tuning is used.In the experimental verification stage,we first conducted an experimental test on the expression 7 classification of the recognition algorithm in this dissertation,and then compared the recognition rates of the facial expressions before and after repair.At the same time,we compared the recognition rates of different algorithms under different occlusion areas,and finally visualized the facial expression recognition experiment further verifies the feasibility and effectiveness of dividing the facial expression recognition problem into facial repair and facial expression recognition.