Researches On Image Representation And Geometric Constraints In Video Coding

Researches on image representation and on applications of geometric constraints in video coding are presented in this paper. Main contributions are as follows: Firstly, an image segmentation method based on Voronoi diagrams is proposed, which utilizes the promising property of Voronoi diagrams, i.e., the shape of a Voronoi region can be determined by the position of its seed, and represents the shape of a Voronoi region by its seed. In addition, an optimization procedure designed in the approach can make edges of a Voronoi region run parallel to image contours. On the basis of the approach, some methods suitable for image content representation within a Voronoi region are further studied and an image representation scheme based on Voronoi diagrams is proposed. Then, triangle oriented shape adaptive DCI is studied and an image coding method based on Delaunay triangulation and adaptive DCT is presented. In this approach, residual signals, which are produced by segmentation and coding based on Delaunay triangulation, are encoded by conducting one-dimensional DCT twice within each triangle. By adaptively adjusting directions of the one dimensional DCT according to statistics of the residual signals, transform efficiency can be further improved. The approach has taken full advantage of image content oriented Delaunay triangulation and transform coding, and has improved the quality of reconstructed images. Finally, geometrical constraints in image sequences, mainly including Epipolar geometry and Trilinear constraint, are studied from the view of video coding. A method to determine mesh node displacement using Epipolar geometry is proposed. By converting the problem of determining mesh node displacement to that of determining corresponding relation between two frames under the constraint of Epipolar geometry, the approach has reduced computational cost and still gives comparatively precise mesh node motion estimation. A motion estimation method based on Trifocal tensor and BMA is also proposed, which gives much better motion compensation results compared with BMA when there are camera translation and scene depth variation.