Research On Restoration Based On Level Set

The 3D reconstruction is one of the main contents of the computer visiontechnique. Since 1990s, the "multiple view geometry" has appeared, in which theself-calibrate and the reconstruction' ideas of the cameras are introduced so that someproblems that can not be solved in the past have excellent results now. But it is stilldifficult in dealing with the stereo matching problems which correspond toestablishing the connections between two views in the multiple views.In fact the difficulties are due to the inverse problem which appears in thecomputer vision. This thesis will talk about a new optimal method which is based onthe principle of variation to solve the matching problem of eyes. Variation methodshave become the general methods dealing with the inverse problems, and have beenintroduced in different regions in Image Processing recently (such as inpainting,denoise, segmentation). In the modern Image Processing pattern, Stereo matching and3D reconstruction is described as a functional relatively some parameter minorproblem based on multiple view geometry theoretics in computer vision. TheEuler-Lagrange equations which are deduced from the variational principle byproviding a set of PDE's. It has been proved that the level set methods, as iterativeregulation methods, are successfulThe thesis will describe the matching of eyes and 3D reconstruction as variationproblems, according to the image equations in the multiple view geometry. The basicideas are introduced, such as how to describe these problems and compute the PDEswith the level set methods. For example, firstly the thesis mainly talk about twohypotheses, one is for the Lambertian hypothesis; about the other hypothesis thatobject plane is parallel to the image plane of camera, and the intensity model is givenwhich is computed on the neighborhood. Then we concern the plane homographyamong the tangent plane to the surface of the object and the two image plane, giving amore accurate model. In this model we loosened two kinds of hypotheses, making thatmodel apply of the scope is wider. This thesis introduces the foundation knowledge of projective geometry andcamera model and symmetric geometry relation of two images at first, and thenintroduced level set in particular, which provides an efficient and robust way ofachieving the surface evolution and to deal automatically with changes in the surfacetopology during the deformation. It also can evolve the initial surface and make it getclose to the object.The detailed form of every method is given, and the 2D experiments based onthe synthetical image are also seen in this thesis.