The Study On Accurate Perception Of High-dimensional Visual Information

With the development of photography technology and computing device,com-puter vision technology has been widely used in various fields like social security,driverless vehicle,medical assistance,precise agriculture and video entertainment.Ac-curate perception and understanding of the environment through computer vision has become an important issue in the age of artificial intelligence(AI).Restoring the three-dimensional structure of the scene by breaking through the traditional imaging,and supplementing multispectral information play an important role in providing the in-put signal for AI.However,due to the complexity of the imaging environment and the inherent defects of the imaging technology,the much noise will be involved when obtaining high-resolution data,which limited the research and application in AI field.This dissertation focuses on accurate 3D reconstruction,hyperspectral image re-construction,and 3D point cloud data classification.Based on the existing signal sparse representation and recovery work,a high-dimensional visual information restoration technique is proposed,which adequately takes advantage of the merits of high-dimensional visual information and low-rank visual restoration algorithm.Tensor representation and sparse measurement of high-dimensional visual information are respectively achieved.The main contributions of the dissertation include:1.We propose a high-dimensional vision recovery method using tensor kernel norm as the sparse metric.As the tensor is good at maintaining structure modeling visual problems,this paper adopts the tensor to represent high-dimensional visual signals and presents the algorithm of precise signal recovery of the sparse metric with the sparse merit as prior,which leads to the excellent performance.2.We propose a novel and robust framework of combining a matrix splitting with multi-view stereo reconstructions to separate reconstruction inaccuracies from a various parameters model for high-accuracy multi-view stereo reconstruction.In-stead of performing the fixed parameters reconstruction procedure,we apply the variational based 3D reconstruction algorithm for multi-times with various param-eters to derive a set of hypothetic 3D models and then synthesized the final result by formulating the problem as a low-rank matrix splitting problem.Benefited from the matrix splitting formulation,the outliers and bad matches,which present as the noise in the synthesized model,are effectively removed and thus leads to a 3D reconstruction with higher accuracy than the existing fixed parameters reconstruc-tions methods.3.We propose a reliable and fast algorithm for automatical classification for tree point cloud using basic geometric features,including the columnar shape of the tree trunk,the plane shape of the ground,and the tiny flat shape of leaves.A new descriptor(local geometric point features and local coordinate)is proposed to describe class characteristics,which is used to distinguish among different classes.The effective-ness of our method is verified by comparing our result and ground truth in the real world.4.We propose a novel model,namely graph-regularized tensor robust principal com-ponent analysis(GTRPCA)for denoising hyperspectral images(HSI).The model benefits from incorporating spectral graph regularization into TRPCA,as it pre-serves the local geometric structures embedded in a high dimensional space.Based on tensor singular value decomposition(t-svd),we introduce a general tensor based ADMM algorithm which can solve the proposed model for denoising HSI.The above methods and algorithm demonstrated high performance on both groundtruth datasets and real world datasets,whose effectiveness is also verified in multi-view stereo,3D point cloud classifications,hyperspectral image denoise.