Font Size: a A A

Face Frontalization Using Generative Adversarial Networks

Posted on:2021-02-06Degree:MasterType:Thesis
Country:ChinaCandidate:C L RongFull Text:PDF
GTID:2428330611466955Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Face frontalization is the process of synthesizing identity-preserving frontal face images from profile face images.Most face images collected by real-time monitoring systems have different postures,which decreases the recognition rates of many methods.Face frontalization methods can address pose variations on face recognition tasks by recovering frontal face images from profile face images.In recent year,many face frontalization methods based on Generative Adversarial Networks(GAN)are proposed.These methods can produce frontal face images without distortion under the large poses.However,the performance of these methods decreases obviously under large poses.Moreover,these methods do not preserve expression information.This thesis proposes two GAN-based face frontalization methods.One can preserve identity information under the large poses? the other can preserve expression information.They are used for face recognition and facial expression recognition,respectively.Both methods contain the generation module and discrimination module.The generation module can recover a frontal face from a profile face,while the discrimination module can encourage the generation module to produce better results.They improve each other by competing against each other in the training process,which can improve the visual quality of synthesized images.The main contributions of our thesis can be described as follows:(1)This thesis proposes a GAN-based face frontalization model which can preserve identity information.It aims to improve the ability to preserve identity under large poses.Different from similar methods,it contains a Feature-Mapping Block and a feature discriminator.Feature-Mapping Block can map the features of profile face images to the frontal space.The feature discriminator can distinguish the features of profile face images with those of ground true frontal face images,which guide the generation module to provide high-quality features of profile faces.They both improve the ability to preserve identity under large poses.(2)This thesis proposes a GAN-based face frontalization model which can preserve ex-pression information.A softmax loss function and a center loss function are used in the training process for preserving the expression.They can map the expression features of the synthesized frontal face to the space of correct expression,which helps to preserve expression information.Experimental results on the Multi PIE,Labeled Faces in the Wild databases(LFW),and Celebrities in Frontal-Profile(CFP)databases show the superiority of our first frontalization method over the state-of-the-art under large poses.Experimental results on the Multi PIE show that our second method obtains high recognition rates on facial expression recognition tasks,which proves that our method can produce expression-preserving frontal face images.
Keywords/Search Tags:Facial Expression Recognition, Face Recognition, Face Frontalization, Generative Adversarial Network
PDF Full Text Request
Related items