| Summary: | 碩士 === 國立中山大學 === 應用數學系 === 87 === In this report, a novel approach for the compression of image sequences is proposed by segmenting the constituent pixels according to the Delaunay tessellation paradigm. A Delaunay tessellation partitions a spatial point patterns into unique non-overlapping triangular areas in the two-dimensional case and tetrahedrons in the three-dimensional one. A pixel in an image sequence is determined by its corresponding frame number and x-y coordinates. Given the ensemble of pixels in a image sequence, the Delaunay tessellation segments the equally-spaced points into tetrahedrons by grouping neighboring pixels with similar intensity enclosed within the same volume.
Each tetrahedron formed is examined for homogeneity by comparing the variance of the tetrahedron with a threshold. A non-homogeneous tetrahedron is split into four smaller tetrahedrons by inserting a new point in the center of gravity, then tests their homogeneity iteratively. The following merge step examines neighboring triangles and groups similar triangles into one larger ones. Each Delaunay tetrahedron formed is characterized by a homogeneous gray level with possible varied region size. The spatial coordinates of the vertices of the tetrahedron reached and the corresponding pixel values are recorded for later decompression. On the decoding side, given the spatial distribution of these vertices, a unique tetrahedron can be constructed by following the paradigm of Delaunay tessellation. The gray levels of the pixels enclosed within the tetrahedron are taken as the average value of the all the intensities of the vertices.
The result of compressing motion image sequences, e.g. Football, is better than MPEG. We have higher PSNR and lower bit rate.
|
|---|
