Study Of Several Critical Issues In Image Synthesis Based On Pixel Transferring

This dissertation presents the study of several critical issues in image synthesis based on the pixel transferring. Particular novelties of this dissertation are the high accuracy estimating and application of the geometrical relationship among the image sequences. The epipolar geometry (bilinearity) of two images and the trilinearity (trifocal tensor) of three images are studied in detail. And with the further consideration, the image rectifying and the matching error model are also studied. The so-called image synthesis means that, with the two reference images of the same scene which captured from different camera view points, some new images from other camera view points could be calculated directly from the two reference images. This means the user could be able to move a virtual camera, so as to generate any other view of the scene, and carry Qut a virtual visit of the scene without the necessary of camera calibrating and 3D scene reconstructing. This technique has numerous applications, among which virtual realities, or high rate video compression. The main contribution of this dissertation includes two aspects. The first one is the new synthetic method based on the two-epipole constraint in estimating the fundamental matrix. With this method, we could get the resolution with the stability of the two epipoles. The second one is the geometrical meaning of the coefficients of the fundamental matrix which are analyzed and developed in this dissertation. The developments of these methods clearly increase the accuracy and stability of the fundamental matrix obtained and have the well understanding geometrical meaning. This dissertation also presents the new method of estimating the trifocal tensor in the union configuration. It has some advantages over the previous methods, such as easy to complete and highness in estimation accuracy. In order to evaluate the algorithms in 3D reconstruction, a model of reconstruction error is also studied. To satisfy the demand of the interactive package for industrial and commercial applications, the image matching error and the camera distortion are taken into account to emulate their efforts in reconstruction. Based on the robust estimation of epipolar geometry and trilinear constraint, the dissertation gives some researches on their applications in the view synthesis. Among these researches, an image dense matching method and the image rectification are implemented. From many simulation tests and comparison with some previous classical methods and some other advanced methods in the relative fields, it is proved that the methods in the dissertation have obviously improved the estimation accuracy and stability of the geometric relationship between image sequence and has the expected results in view synthesis.