Research On Filtering And Registration Algorithm Of Three-dimensional Image

Three-dimension reconstruction is an important part in reverse project, and is also a significant research area in computer vision field. In the process of three-dimension reconstruction, unreasonable noise inevitably exists in the three-dimension data of the object surface. Meanwhile, restricted by the measuring angle, the three-dimension data of the object surface needs to be got from different perspectives. Hence, the filtering and registration of three-dimension image are especially significant. This paper studies filtering and registration algorithm of three-dimension image and the related issues.In view of filtering of three-dimension image, the paper firstly studies Laplacian, Taubin and Curvature filtering algorithm. Then, considering the influence degree of the distance between two points, the paper proposes an improved weighted Laplacian filtering algorithm that can keep the original image feature information to some degree and inhibit the image deformation. At the same time, using the normal vectors and the angle between them and introducing Laplacian operator as deviation coefficient, this paper proposes an improved median filtering algorithm on normal vectors as a foundation,which can keep the image feature information and get better filtering effect. Finally, the paper compares and analyzes the test results of several filtering algorithms.Aiming at registration issue of three-dimension image, this paper firstly studies ICP registration algorithm, and then analyzes each stage of the algorithm and selects the methods at different stages through experiment. To solve the problems that the number of data points is too large to compute and the characteristic of selected points isn't obvious, the paper proposes an improved ICP algorithm based on feature extraction. Firstly, it extracts the key points of three-dimension image. Afterwards, it selects corresponding points according to the similarity of their feature descriptors and estimates the rigid body transformation matrix for registration using Quaternion. At last, this paper analyzes and compares the experimental results of different algorithms. The results shows that the improved ICP algorithm has faster convergence speed, better registration effect and also has good stability.