Sparse Three-dimensional Reconstruction For Big Data Image

Facing the massive,unstructured and non-precise characteristics of big data image,the traditional photogrammetric process has encountered great challenges in terms of applicability,computational efficiency and completeness of results.Based on the above background,this dissertation mainly studies how to recover three-dimensional information from big data image,including the analysis of image correspondence based on feature points,the pluralistic description and the cluster analysis of big data image,the pose optimization with nonlinear optimization,the sparse three-dimensional reconstruction of big data image and other aspects.The dissertation effectively solves the big data image processing in the calculation of the timeliness,the results of integrity and other issues,promotes the development of big data image mining analysis.In general,the main work and innovation of this paper are as follows:1)A strategy of block extraction and packet matching based on GPU is designed,which provides sufficient information of matching points for clustering and reconstruction.The experimental results show that the strategy can make full use of the hardware performance of the computing platform,and quickly extract rich matching points.2)A matching points purification model based on principal component analysis is proposed.The model uses the whole matching points as initial input,and gradually removes the wrong points to get more accurate global optimal solutions.The experimental results show that the proposed method can obtain a global optimal solution under certain original misplaced ratio,and decrease the omission of correct matching points at the same time.3)A clustering algorithm for massive and disordered images is designed.Firstly,combined with the edge feature of the image,a kind of aggregation descriptor,which can express the whole semantic and detail characteristics of the image,is proposed to carry out the global description of the image.In the descriptor clustering stage,considering the time and accuracy requirements,a real-time self-organizing feature map neural network with simple parameters and without prior training is proposed to perform the final clustering.The experimental results show that the algorithm can perform real-time clustering of big data image,has high accuracy and good stability.4)An adjustment model based on preprocessing conjugate gradient is introduced to solve the problem of the three-dimensional reconstruction.The experimental results show that the optimization algorithm can quickly and robustly calculate massive data to complete the pose optimization task.5)A sparse three-dimensional reconstruction method for big data image is proposed.Combined with the modular big data image clustering method,a two-layer incremental sparse three-dimensional reconstruction method is adopted to realize the fast scene sparse three-dimensional reconstruction.The experimental results show that the method can quickly analyze big data image with low hardware requirements,and obtain more complete sparse three-dimensional information.