Font Size: a A A

Pose Normalization Algorithm Based On 3D Face Reconstruction And Image Inpainting

Posted on:2021-05-08Degree:MasterType:Thesis
Country:ChinaCandidate:R S ZhengFull Text:PDF
GTID:2428330614960416Subject:Electronic and communication engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of artificial intelligence and computer vision technology,face recognition technology has got extensive attention of researches all over the world.At present,there are many mature face recognition systems has been used in practice.But in the actual scenario,there exists some insufficiencies for face recognition systems.One of the most important factors is face pose variation that always restricts the performance of face recognition system.Under this background,in this paper,a face pose correction algorithm based on 3D face reconstruction and image inpainting is studied.The main work of this paper is summarized as follows:(1)In this paper,we introduce the study history and the mainstream algorithms of face recognition and analyze the components of the face recognition system and the factors affecting the performance of face recognition.Especially,we emphasize the limitation of face recognition rate due to face pose variation and give a detail introduction of researches of multi-pose face recognition all over the world.In addition,we also give a detail introduction of researches of 3D face reconstruction all over the world and highlight the role of 3D morphable model in face image reconstruction.In the end,we introduce the basic information of some common face datasets.(2)In order to solve the problem of low face recognition rate due to face pose variation,in this paper,we propose a profile face image reconstruction algorithm by using 3D morphable model.The details are as follows: we make it possible to reconstruct the face image under the large face pose by updating the landmarks on the contour of 3D face model.In addition,we assign different weights to the landmarks on the different regions of face,and reconstruct face image by the weighted sum of these landmarks.It makes the reconstruction effect more robust to the facial expression changes.In the end,we can see the effectiveness of our proposed face image reconstruction algorithm from the experiments on the Labeled Faces in the Wild face dataset and Stirling ESRC 3D face dataset.(3)We will get face image with invisible region after pose correction for 3D face model due to the self-occlusion caused by face pose variation.The performance of face recognition system will decrease when recognizing the face image with invisible region.In order to solve this problem,we utilize the method of image inpainting.Firstly,we give a deeply introduction of study history of image inpainting algorithms and propose two face image inpainting algorithms: face image inpainting algorithm based on lambert illumination model and face image inpainting algorithm based on deep learning.Then we use these two methods to repair the face images with invisible region.The importance of image inpainting algorithms will show from the qualitative and quantitative experiments on the Labeled Faces in the Wild face dataset and Stirling ESRC 3D face dataset.
Keywords/Search Tags:multi-pose face recognition, 3D face reconstruction, image inpainting, deep learning
PDF Full Text Request
Related items