Research On Joint Depth Map Completion And Denoising Technology Based On Graph

The depth image represents the distance of the object from the camera in the scene,which has been widely used in the fields of 3D reconstruction,refocusing,robot vision and so on.The most typical method for obtaining the depth map is based on the method of structured light.However,due to the limitation of the depth sensor,the obtained depth map often has degradation such as noise and edge loss.The degraded image will greatly affect the subsequent use,so the depth Image restoration is a very important topic.In this paper,we propose two depth map repair methods based on graph structure.Because modern depth sensors are always accompanied by natural image sensors,and the acquired natural images are basically not degraded,most of the deep repair methods are based on color image-guided repair techniques.Depth image restoration is essentially a morbid problem,and it requires a certain prior knowledge to guide.Commonly used a priori have low rank and total variation based on color image guidance.Aiming at the a priori deficiencies of existing matrix-based decomposition-based degenerate low rank matrix repair methods,this paper introduces a dual-graph Laplacian regularization based on local pixel smoothing and non-local gradient consistency by introducing local and non-local priors.With dual-graph regularization,the objective function can achieve fast repair of the depth image.By performing block-matching on the depth map after pre-complementing with NLM,a low rank matrix is ??obtained,thereby converting the image restoration problem into a low rank matrix repair problem.We treat each row of the low-rank matrix as the vertex of the graph,Then we construct a local graph regularization representing local pixel similarity.We construct a column graph regularization which representing gradient consistency with each column of lowrank matrix.The objective function constrained by the dual graph has a closed-form solution,thus we can achieve fast repair of the depth map.Finally,the restored low-rank matrix is ??restored to the corresponding position in the depth map to obtain a complete repair depth image.Experiments show that the local and non-local constraints in the dual-graph achieve complementary effects.For the problem that the low-rank prior is not used in the problem of low-rank matrix repair,this paper proposed to put nuclear norm,which is the convex relaxation of the lowrank prior,into the regularization with dual-graph.The objective function is solved by ADMM.When the ADMM sub-form obtained by the decomposition is actually solved,the optimization is made to make the three sub-steps have a closed-form solution.The experimental results show that the repair results using the low rank and dual graph jointly exceed the state-of-art in PSNR and SSIM,and the optimized ADMM subroutine calculation is also faster.