Lossless Contour Compression Using Chain-code Representations and Context Tree Coding

In this paper we experimentally study the context tree coding of contours in two alternative chain-code representations. The contours of the image regions are intersecting, resulting in many contour segments, each contour segment being represented by its starting position and a subsequence of chain-code symbols. A single stream of symbols is obtained, under a given ordering of the segments, by concatenating their subsequences. We present efficient ways for the challenging task of ordering the contour segments. The necessary starting positions and the chain-code symbols are encoded using context tree coding. The coding is performed in a semi-adaptive way, designing at the encoder by dynamical programming the context tree, whose structure is encoded as side information. The symbols are encoded using adaptively collected distributions at each context of the encoded tree. We study for depth map images and fractal images the tree-depth of the optimal context tree and the similarity of the context trees resulted to be optimal for each image.