Research On Single-depth Image Inpainting Algorithm

Depth images are an important representation of depth information and are a research hotspot in the field of image processing and computer vision.Depth images enable computers to obtain three-dimensional information through two-dimensional plane images,which contributes to a better expression of the three-dimensional structure of the scene and promotes the rapid development of machine automation,target tracking,medical diagnosis and treatment and other fields.However,due to the limitation of many factors,it is easy to cause the lack of depth image information.Therefore,it is of great theoretical and practical significance to use a reasonable algorithm to repair to obtain high-quality depth images.Single-depth image inpainting is an ill-posed inverse problem.The prior depth image can adopt the appropriate mathematical language as the constraint condition to change the inpainting problem from ill-posed to well-posed to complete the repair.In this thesis,based on the characteristics of low rank structure,non-local self-similarity,sparsity and gradient sparsity of the depth image,the problem of single depth image inpainting was systematically and deeply studied.On the basis of previous work,three different inpainting methods were proposed,and the task of single depth image inpainting was completed.The main innovation points and specific work are as follows:（1）A single-depth inpainting method based on nonlocal self-similar features and low-rank constrained is proposed.The depth image is a description of the geometric shape of objects in the scene,and the edge structure between objects is important.In this method,a low-rank representation model is used to obtain a rough restoration effect.On this basis,nonlocal self-similar constraints are introduced to improve the inpainting effect because the depth image has a large range of regions with the same depth value and repeated parts of the edge structure.This method can make full use of the features of depth images to complete the inpainting task.First,according to the characteristics of the pixel value,the depth image is partitioned to form similar block groups and a three-dimensional arrangement.Then,the inpainting problem is effectively divided into low-rank constraint subproblems and nonlocal self-similar constraint subproblems by using the split Bregman iterative algorithm.Finally,different strategies are used to solve different subproblems,which improves reliability.Weighted Schatten P-norm minimization as a low-rank constraint solution strategy can not only adopt different operations for different parts but also approach the original low-order hypothesis better and obtain a potentially intact solution.The pixel block-level nonlocal self-similar sparse statistical model in the 3D transform domain can better describe the self-similar degree of images and better represent the real signals in the group.（2）A single-depth image inpainting method based on gradient features and sparse representation is proposed.The depth image can be regarded as a natural image without texture,which is composed of the most similar flat areas and a few edge areas.On the one hand,different from the traditional sparse representation of image blocks,the sparse model based on similar block groups can uniformly represent the local smoothing features and nonlocal self-similar features of depth images.On the other hand,the gradient L₀norm smooths the unimportant details in the depth image by constraining the number of nonzero gradients to improve the edge salience.As a method to solve the L₀norm of the gradient,the region fusion criterion can better describe the features without texture.The method makes full use of the characteristics of depth images to achieve inpainting.（3）A single-depth image inpainting method based on ELU-CNN network is proposed.With the help of the split Bregman algorithm,the depth image inpainting problem is transformed into a denoising subproblem,and then a denoising network based on residual learning is introduced.This model adopts the dilated convolution and ELU activation function,combines a residual network with batch normalization,and trains a group of efficient convolutional neural network denoising piror to solve the subproblems of the depth image and complete the task of single-depth image inpainting.Dilated convolution can balance receptive field and the depth of network.ELU activation function can reduce the offset and improve robustness to noise.The combination of a residual strategy and batch normalization can enhance the stability of the training.