作者: Shu Zhang , Ran He , Zhenan Sun , Tieniu Tan
DOI: 10.1109/TIFS.2017.2763119
关键词:
摘要: MeshFace photos have been widely used in many Chinese business organizations to protect ID face from being misused. The occlusions incurred by random meshes severely degenerate the performance of verification systems, which raises problem between and daily photos. Previous methods cast this as a typical low-level vision problem, i.e., blind inpainting. They recover perceptually pleasing clear MeshFaces enforcing pixel level similarity recovered images ground-truth then perform on them. Essentially, is conducted compact feature space rather than image space. Therefore, paper argues that jointly offer key improve performance. Based insight, we novel oriented inpainting framework. Specifically, implement establishing DeMeshNet, consists three parts. first part addresses implicitly exploiting extra supervision occlusion position enforce similarity. second explicitly enforces space, can explore informative produce better results for verification. last copes with alignment within net via customized spatial transformer module when extracting deep facial features. All parts are implemented an end-to-end network facilitates efficient optimization. Extensive experiments two data sets demonstrate effectiveness proposed DeMeshNet well insight paper.