Research On Occluded Face Inpainting

As a branch in the field of image restoration,face inpainting technology has the significant application value in many scenarios.For example,in the security video surveillance system,if the acquired face image is partially occluded,the key features of the face will be highly destroyed,and the more accurate inpainting of the occluded area is a major breakthrough in improving face recognition technology.Meanwhile,face completion is also an essential operation in image editing,which requires generating harmonious content and visually realistic image information to fill the masked area.At present,due to its unique geometric structure,the face has some challenges in its occlusion generation,and it is easy to brings about producing textures with blurred borders and unreasonable context perception.This paper conducts research on this issue and constructs the face inpainting model based on convolutional neural networks.Considering that the existing methods have not fully explored the prior effect of face geometry during the completing process,this paper proposes a two-stage face inpainting framework:In phase one,based on the generation of adversarial networks,a geometric structure inference model of the face is proposed in the early stage of completion,and capturing the facial parsing maps emerged each component's semantic presentation through estimating the prior information of the occluded face;In phase two,combining the parsing map obtained in phase one,proposes a three-input generative inpainting model to complete the reconstruction of the texture details of the occluded area,that is,based on the partial convolutional network,establishing encoders with the masked image(appearance),parsing map and binary mask separately,extracting depth features at different levels,and fuse appearance features and parsing features in each decoding stage to play the guided-inpainting function of geometric priori.In order to further improve the generating effect of the model,when merging the parsing features,this paper takes into account the correlation between the occluded area and the global face structure and the degree of fusion of important parts.Therefore,convolution block attention module is introduced after the middle and high-level features concatenating during the decoding stage.The block attention mechanism,from the perspective of feature channels and pixels spacial position,enhancing the expressiveness of features,reducing the interference that may be caused by the lack of local background,and generating semantically coherent face image details.The experiental results on the CelebA standard dataset verify the effectiveness of the proposed algorithm.