Research On Image Inpainting Algorithm Based On Generative Adversarial Network

Image inpainting technology has always been a research hotspot in the field of image processing.With the rapid development of this technology,it has been gradually applied to many fields such as cultural relic restoration,medical imaging,satellite remote sensing,etc.Therefore,the research on image inpainting algorithms has high research significance and application value.The purpose of image inpainting is to use computer technology to independently learn the characteristic information of the image,so as to restore the content of the missing area of the image to be repaired as much as possible.Compared with traditional image inpainting algorithms,image inpainting algorithms based on deep learning usually have better repair effects.However,there are still many problems,for example,poor repair effect for images missing in large or irregular regions,unclear repair texture,and inability to use auxiliary information to guide the repair result.In addition,since the Generative Adversarial Network can achieve the purpose of generating a sample that is close to real through the adversarial learning between the generator and the discriminator,this idea is suitable for solving the image inpainting problem.Therefore,this paper proposes two image inpainting algorithms based on the research of Generative Adversarial Networks.The specific research contents are as follows.This article first introduces the definition and evaluation criteria of image inpainting.Then the related principles of convolutional neural networks are introduced,and the shortcomings of dilated convolution and deconvolution are analyzed in detail.Finally,the principle of image inpainting based on Generative Adversarial Networks is studied and analyzed,which provides a theoretical basis for the subsequent research and improvement of image inpainting algorithms in this paper.Aiming at the problems of structural distortion and texture blur in the image restoration results of missing large or irregular regions,this paper proposes an image inpainting algorithm based on multi-scale texture feature branches.Firstly,the advantages of the evennumbered convolution kernel with symmetrical padding are analyzed,and uses the evennumbered convolution to construct the overall network model of the algorithm.Then in the generator stage,combined the partial convolution network and the even convolution to design the multi-scale texture feature branch to realize the simultaneous restoration of image structure features and texture features by a single-stage network model.Meanwhile,interpolation is used to avoid the checkerboard effect caused by deconvolution upsampling;In the discriminator stage,the dual discrimination structure is used to ensure the consistency of the overall structure of the repair results and the rationality of the local details,and the idea of spectral normalization is introduced to stabilize the network training.Finally,experiments prove that the algorithm can obtain higher quality repair results.For some image restoration problems that can add additional auxiliary information,in order to use auxiliary information to guide the image inpainting process,this paper proposes an image inpainting algorithm with gating and contextual attention.According to the update method of the gated convolution self-learning mask,the algorithm uses a gated convolution network to achieve the overall construction of the generator model to allow processing of additional auxiliary information input by the user;then introduce the receptive field module and loss function to improve the context attention mechanism,and the improved attention module is embedded in the encoder stage to enhance the utilization of the texture features of the known region.Finally,it is verified through experiments that the algorithm can guarantee the superiority and effectiveness of the repair result without adding the guidance information,and the algorithm can also obtain a higher restoration degree under the premise of ensuring the quality of the repair when the guidance information is added.