Research And Implementation Of Image Completion Based On Generation Adversarial Networks

Image completion is an important research direction in the field of digital image processing,and has a broad application prospect.In recent years,the generation adversarial network(GAN)provides a new method and technology for image processing tasks.Compared with the traditional deep learning,it has a stronger ability of expression and feature learning,and has achieved remarkable success in the field of image completion.However,the research of image completion using GAN mainly focuses on the restoration of the missing content of the image,and the feature of the generation model is to generate new information.Based on this feature,mark Sabini and others proposed the concept of image "outpainting".Image "outpainting" goes further on the basis of image inpainting.It can "predict" the part outside the boundary that does not exist in the original image according to the image fragment,that is,given an original image,an image with greater resolution and more information can be generated through image outpainting.Image outpainting will generate many exciting and novel applications.For example,we can use the image output for panoramic creation,vertical video extension and texture creation,but there are few related researches,so it has great research significance.Image inpainting and image outpainting are closely related,but there is no clear research that the former technology can be directly applied to the latter.In this paper,the leading edge technology of image inpainting is introduced into image outpainting:(1)The residual block is introduced to improve the convergence rate of the model and avoid the gradient vanishing / gradient explosion;(2)We use Wasserstein distance to train the model,improve the stability of the model,and generate the diversity of pictures.In particular,we do not use a single discriminator,Two discriminators are used in this paper: the global discriminator network and the local discriminator network.As we have proved,this is very important for obtaining the semantic and local coherent image completion results;(3)In order to further improve the image completion performance of the generated network,the energy function of MRF is introduced while MSE loss training is added to the generated network.By adding this energy function,the consistency between the missing pixels and the edge pixels in image features is guaranteed.It enhances the consistency of image features between the missing part and the whole part,at the same time,the loss function decreases faster,and improves the image effect of the generated network again.Finally,we evaluate the performance of the extended image by training convergence rate,structure similarity,peak signal-to-noise ratio,normalized mean square error and average absolute error.After training the model on the beach and grassland data set separated by place365-Standard,the more realistic image is completed,which proves that we have successfully improved the model,our model loss convergence is faster,and the quality,authenticity and diversity of the completed image are improved.In addition,we also designed four groups of modified super parameters,which are gradient penalty parameter,CT penalty parameter,Markov random field weight parameter and traditional GAN adversarial loss weight parameter in joint loss,to obtain the global optimal state of the improved model.