标题:A deep variational autoencoder approach for robust facial symmetrization
作者:Wang, Ting ;Zhang, Shu ;Dong, Junyu ;Liang, Yongquan
通讯作者:Zhang, Shu
作者机构:[Wang, Ting ;Liang, Yongquan ] Shandong University of Science and Technology, Qingdao, China;[Zhang, Shu ;Dong, Junyu ] Ocean University of China, Qin 更多
会议名称:29th British Machine Vision Conference, BMVC 2018
会议日期:3 September 2018 through 6 September 2018
来源:British Machine Vision Conference 2018, BMVC 2018
出版年:2019
关键词:Facial symmetry; Multi-stage training; Variational autoencoder
摘要:Face symmetrization has extended applications in both academic and medical fields. Human face possesses an important characteristic, which is as known as symmetry. However, in practice, this symmetry is never perfect, which yields a large amount of studies around this topic. For example, facial paralysis evaluation based on facial asymmetry analysis, facial beauty evaluation based on facial symmetry analysis, facial recognition, and facial frontalisation among others. Currently, there are still very limited researches that are dedicated for this topic. Most of the existing studies only utilized their own implantations for symmetric face generating to achieve their researches in other fields. Thus, limitations can be noticed in their methods, such as manual intervention requirement. Furthermore, most existing method utilize facial landmark detection algorithms for symmetric face construction. Despite the promising accuracy of the landmark detection algorithms, the uncontrolled conditions in facial images can badly impact the performance of the symmetric face production. To this end, this paper presents a variational autoencoder based deep generative model for symmetric face generating. It is achieved by a 3-stage training process to avoid the demand for the large size of the symmetric face as training data. Experiments are conducted with comparisons with several methods that achieved by some of the most popular facial landmark detection algorithms. Competitive results are achieved. © 2018. The copyright of this document resides with its authors.
收录类别:EI;SCOPUS
资源类型:会议论文;期刊论文
原文链接:https://www.scopus.com/inward/record.uri?eid=2-s2.0-85072346514&partnerID=40&md5=de46475c6407518d37bf342aa0e67a72
TOP