The Face Inpainting Based On Edge Priori And Non-local Network

Deep learning has been widely used in image classification,target detection,image segmentation,speech recognition and other fields in recent years.Image inpainting is an important task in the field of computer vision.It is common in our daily life,such as image and video production,medical imaging,public safety,and cultural relic restoration.Image inpainting is to use the structure and texture of background to fill harmonious and consistent content for the missing region.Deep learning improves learning ability and refines results of inpainting algorithms,which commendably solves the limitations of traditional methods.Today,plenty of inpainting methods for different problems have been proposed.How to further utilize the background and texture of image for completing image has become the focus of this article.This paper analyzes the development of image inpainting and introduces some existing cutting-edge technologies.Furthermore,the application of deep learning and other important technologies in image inpainting is introduced,as well as some problems still faced in image inpainting.An edge completion network based on feature fusion is proposed.When extracting the features of edge information,low-level feature maps and high-level feature maps are concatenated and fused to provide more semantics such as lines,points,and corners for completing the edge contour.Therefore,it not only hallucinates a more accurate and refined edge structure,but also provides a more reasonable prior information for color image inpainting.Inspired by non-local neural network,we add a non-local convolution block in the process of down-sampling and residual convolution of the image generation network,which is to ensure the coordination of color distribution and the consistency of semantic structure of the reconstructed image.By learning feature information of a certain region and learning longrange information at the same time,the receptive field can be enlarged and the global feature information can be more fully utilized.In this paper,face image datasets are trained,and some algorithms are compared qualitatively and quantitatively through simulation experiments.The experimental data show that the proposed model is superior to some classical algorithms.