Research And Application Of Image Inpainting Technology Based On Generative Adversarial Network

Image restoration has a long history as an important part of the field of digital image processing.From early cultural relic restoration to current intelligent security,cultural relic mural restoration,remote sensing satellite,medical image enhancement and other fields,image restoration has a wide range of applications.With the research and development of digital image restoration technology,image restoration methods can be roughly divided into three types:traditional variational partial differential equation methods,methods using texture blocks,and deep learning methods.Traditional methods can produce good repair results in some fields,but there are also problems that the repair results are relatively fuzzy and can only be used for images with small missing areas,and cannot repair missing areas of any size.At present,deep learning related technologies are widely used in the field of image restoration.For damaged images with different types of missing missing sizes and different missing sizes,they have a good improvement in the details and quality of the repair,such as the classic codec image repair model.Based on the two-stage encoding and decoding network,this paper proposes an improved two-stage image repair model with multi-scale structure embedding and attention mechanism.Then,the improved two-stage image repair model has more parameters and deeper layers,and the training phase generates Improve against the problem of network instability.First of all,this paper proposes an improved two-stage image restoration model based on the two-stage encoding and decoding restoration model.The generator is composed of a twostage repair network from coarse repair to fine repair.In the coarse repair stage,expanded convolution is used to increase the field of perception during the convolution process;in the deconvolution reconstruction process,more detailed information is generated in order to obtain more structural information.The texture boundary embeds multi-scale structural information.In the fine repair stage,the upper and lower parallel network structure is used to optimize the rough repair results.On the one hand,the contextual attention mechanism is used to focus on repairing the surrounding area and better fit the texture and color information;on the other hand,the expanded convolution is used to increase the receptive field and borrow more peripheral information to improve the quality of damage repair.In the training process,in order to make the model more stable and easy to control,the WGAN-GP discriminator is used.The discriminator is composed of a local WGAN-GP discriminator and a global discriminator WGAN-GP.The local discriminator is used to judge whether the repair result is consistent with the surroundings,and to improve the quality of the repair boundary.The global discriminator is used to judge whether the restored image is consistent with the original image,and improve the consistency and fluency of the overall color structure of the restored image.In order to evaluate the pros and cons of the model,a combination of subjective evaluation and objective evaluation is used to verify the repair effect of the proposed improved model.The subjective evaluation verifies the superiority of the model restoration effect by visually comparing with the images restored by the three models of GL,CA,and EG.Objectively evaluate different repair models GL,CA,EG and the method in this paper on the same data set,and compare and verify the three indicators of PSNR,SSIM and average L1 loss.Secondly,although the improved two-stage image repair model can produce a good repair effect on damaged images,due to the sparseness of the network,the model structure is relatively deep,resulting in weakened or even forgotten structure information,which appears when fusing structure and content information Distortion and blurring,so it produces a poor repair effect.Aiming at the problems existing in the two-stage image restoration model,an image restoration model based on structure embedding is proposed.In the generation stage,a multi-task framework is used to reconstruct the extracted structural information in the process of image deconvolution.At the same time,it uses a multi-scale attention mechanism,uses image channels of different scales,and uses neural networks to make full use of the extracted information to help generate repaired images with clearer texture and more suitable structure.In the identification stage,in order to improve training stability and speed up convergence,Patch-GAN discriminator and pre-trained fixed VGG-16 network are used for structure generation learning and image generation learning.In order to obtain high-level structural information and ensure the consistency of the restored image style,perceptual loss and style loss are introduced in the training process to improve the quality of the restored image.Through ablation comparison,analyze the role of different modules in the improved model in the repair process,and verify the advantages and disadvantages of the improved model with other models under subjective and objective evaluation methods.