Hybrid Scheme with Triangulations for Transform Coding

Abstract

Our approach to image region approximation offers a complete scheme consisting of several steps. We separate the encoding of region boundaries from the texture description. The original image is first segmented using an unsupervised segmentation method for color-texture regions. Following polygonal approximation of created regions causes the degradation of region boundaries. The triangulation is then applied to polygons and either all short edges, or all small triangles are filtered out from the triangular mesh. This results in new smaller regions. The encoding and decoding of polygons and triangles is very efficient, because we need to store only the vertices. For texture approximation we use 2D shape independent orthogonal transforms (e.g. DCT II). The texture is encoded with a code similar to JPEG arithmetic code. The encoding scheme proposed in this paper is much faster than latest approaches with polygonal approximation. We present the two triangulation algorithms - constrained greedy (CGT) and constrained Delaunay triangulation (CDT). Both CGT and CDT provide similar results.