Research On Multi-dimensional Image Inpainting Based On Deep Learning

The image inpainting task aims to use computers to generate natural and reasonable fills for the missing areas of an image.When extending image inpainting to the 3D dimension,the task becomes 3D image inpainting.After a single image is projected into 3D space,due to the occlusion relationship between the foreground and background,there will be missing content near the foreground edge,so it is difficult to restore the complete 3D scene.The task of 3D image inpainting is to restore the conversion effect between two perspectives by filling in missing 3D scene information.This collection of multiple perspectives is a 3D image,and the3 D image will have realistic visual effects after rendering into a video.The 2D image inpainting and 3D image inpainting tasks are collectively referred to as multi-dimensional image inpainting.Although multi-dimensional image inpainting integrates a variety of computer vision technologies to optimize the build details,in the face of lighting,occlusion and other problems in practical tasks,there are still problems of incomplete repair and repair errors.Therefore,how to improve the performance of multi-dimensional image inpainting tasks is the main research goal of this paper.Aiming at the problems of misalignment of local details and low inpainting accuracy in2 D image inpainting,an image inpainting algorithm based on edge completion is proposed in this paper.The algorithm uses edge information as prior to strengthen the comprehensive processing capability of the network,and uses two adversarial generative networks to inpaint the edges and missing areas of the image respectively to generate high-precision image inpainting results.First,the edge completion network integrates scene content and edge information to generate object edges in missing regions.After that,the image inpainting network uses the completed edge information as a prior to guide the inpainting of image content,thereby generating natural and continuous inpainted images.The algorithm uses dilated convolution,a better discriminator,and a composite loss function to improve the network’s ability to perceive local information,thereby generating high-quality inpainting results.Experimental results show that by adding edge prior information,the proposed algorithm can generate reasonable repair results for complex defect regions,and demonstrate superior generalization ability in different scenarios.On the basis of 2D image inpainting,this paper proposes a multi-stage optimized 3D image inpainting algorithm.This algorithm adopts the idea of multi-stage algorithm,including three stages: depth acquisition,depth processing,and 3D generation.First,in order to extract more complete depth scene information,the algorithm introduces a neural network structure search technology to the depth estimation to obtain a network structure with better performance.Secondly,the algorithm adopts a variety of data enhancement strategies,and designs a variety of image post-processing optimization algorithms for different practical cases to enhance the effect of depth scene information in constructing 3D scenes.After that,the algorithm also added an image segmentation network to assist scene classification,and layered images according to the subject of the scene,thereby improving the efficiency of inpainting.In addition,the algorithm sets different camera movement modes for different scenes to improve the authenticity of the reconstructed scene.The experimental results show that the algorithm can accurately repair the 3D scene in a relatively short time,and has a robust processing ability for a variety of actual scenes.