Research On Image Completion Algorithm Of Generative Adversarial Network Based On Image Edge Structure

With the further development of 5g technology,the use threshold of the Internet of things is further reduced,which enables it to be applied in more fields,such as face recognition of people passing by.This also extends the needs of public security and society,and there is a demand for identity recognition that blocks faces.However,the existing face occlusion restoration methods still have the problem of poor visual connectivity of facial features and other details.This paper proposes an image edge completion method including global discriminator network,local discriminator network and face part discriminator network.The role of global discriminator network ensures the visual connectivity of the whole picture;The local discriminator network ensures the visual connectivity of the filling part;The facial part discriminator network ensures that the facial features comply with semantics.The global discriminator network,local discriminator network and face part discriminator network are combined to form an image complement discriminator network.Input the incomplete gray image,incomplete edge image and incomplete mask to the edge completion network,output a repaired edge image,and then input the repaired edge image and defect image to the image completion network to generate a repaired complete face image,and finally input the defect image for testing.On this basis,we have done the following work and achieved certain results:(1)In the image completion network,a multiple discriminator is proposed.It is proposed to use the local discriminator and the face part discriminator,use the local discriminator to evaluate the overall effect of the completion part,and use the face part discriminator to evaluate the effect of the face’s eyes,nose and mouth.This method effectively improves the information reconstruction effect of the face part,and also pays attention to the relationship between the whole and the part,which makes the reconstructed information coordinated with the style,texture and edge structure of the whole image,so as to achieve the effect of more accurately restoring the original facial features.(2)The residual block in the image completion network is replaced by the self attention module,and the gated convolution and incentive mechanism are added to the self attention module to form a self attention module combining the gated convolution and incentive mechanism.This module enables the image completion network to more accurately distinguish the characteristics of the complete part and the missing part from the boundary of the mask part,and solve the problem of the limitation of the receptive field of convolution.This enables the mask part to restore a part that fits the texture through the whole information and the whole part side of the boundary at the same time.