Research On Image Inpainting Algorithm Based On Deep Learning

As an important carrier of human information transmission and preservation,images are widely used in satellite remote sensing,information science,medical imaging and many other fields.However,in the process of creation and preservation of image information,the image is often affected by unexpected situations,resulting in image damage and affecting subsequent research.Therefore,how to restore image information to the greatest extent and reduce the information loss caused by image damage has become a hot issue for researchers.Image inpainting aims to predict the content of the missing area according to the known prior information of the image,and finally obtain a reasonable inpainting result consistent with the overall structure of the image.Most of the existing image inpainting algorithms are based on deep learning.Among them,the image inpainting algorithm based on Generative Adversarial Network is widely used because of its good performance,but there are still many problems,especially in repairing large-area missing images and irregular damaged images,such as distorted structure,fuzzy details,uneven color,and unnatural edge transition.To solve these problems,this thesis uses the Coarse-to-Fine network of the Generative Adversarial Network as the basic structure,and proposes two image inpainting algorithms.The details are as follows.In order to solve the problem that existing image inpainting algorithms are not effective for large-area missing images,this thesis proposes an image inpainting algorithm based on multi-scale hybrid attention mechanisms.In the generator stage,a Self-Calibrated attention module is constructed combined with residual connections.And this module and spatial SelfAttention module are used in the fine network to improve the network’s ability to analyze global semantics and local details.In the discriminator stage,the global discriminator and the local discriminator are used to discriminate the inpainting results from both the global and local aspects,and combine various loss functions to update the network parameters,so as to guide the generator to generate the inpainting results that are indistinguishable from the real.The existing image repair algorithms often have some problems such as uneven color and image distortion when repairing irregularly damaged images.In order to solve these problems,this thesis proposes an irregular damaged image inpainting algorithm based on Gated convolution.In the fine network stage of the generator,a hole residual module is used to expand the network receptive field to ensure full utilization of effective pixels.In the discriminator stage,SN-Patch GAN,which is more suitable for discriminating irregular damaged images,is introduced to stabilize the training process of the network.At the same time,gated convolution is used to replace ordinary convolution in the network,which allows the network to automatically learn masks and effectively improves the inpainting quality of irregular damaged images.This thesis verifies the effectiveness and superiority of the two proposed algorithms in solving image inpainting problems through multiple sets of comparative experiments.