[extensive reading of papers] attributes guided facial image completion

Thesis title ：Attributes guided facial image completion
Time ：2020
source ：Neurocomputing
Address of thesis ：https://doi.org/10.1016/j.neucom.2020.02.013
There is no public code in this paper
There is no public code in this paper
There is no public code in this paper

Abstract
   The face image completion method reproduces the missing pixels of the face image , Make it difficult to distinguish the completion result from the original image . Multimodality of image content , There are usually many specious ways to complete facial images . Users may be particularly interested in obtaining a specific completion result with certain required attributes . To meet this demand , We propose a new deep learning method to complete the task with multiple controllable attributes ( Such as gender and smile ) Face image of . First , An attribute classification loss algorithm is proposed , To ensure that the finished image has the required properties . secondly , Our method introduces perceptual loss to model the high-level semantic structure in face images , This greatly reduces structural distortion in large missing areas . Third , Use adversarial losses , To ensure that the finished results are visually authentic . These three components fit well in our approach , Formed an effective framework , High quality face image with multiple controllable attributes . stay CelebA Experimental results on data sets show that , This method not only achieves good performance in face completion task , It also allows users to obtain specific completion results according to the face attributes they are interested in .

   In recent years , Due to the deep convolution network, especially Generative adversary network Rapid development of , There are many methods based on deep learning , Good results have been achieved in the task of image completion . These methods adopt the current popular convolutional codec network structure , And have been trained to fight against losses , Select a reasonable completion mode from many candidate completion modes , It not only synthesizes low-level texture information , New high-level semantic structure information is also generated . Especially for facial images , These works have proved to be effective in generating credible new content .

   Unfortunately , These methods do not control the properties of the composite content , Only focus on fitting according to the known context , To make the finished image look natural and real visually . However , Users may be particularly interested in using certain required attributes ( for example , Smiling or not smiling ) To complete the missing area .

Generate face images with controllable attributes