Coding theoretic approach to image segmentation

This paper introduces multi-scale tree-based approaches to image segmentation, using Rissanen's coding theoretic minimum description length (MDL) principle to penalize overly complex segmentations. Images are modelled as Gaussian random fields of independent pixels, with piecewise constant mean and variance. This model captures variations in both intensity (mean value) and texture (variance). Segmentation thus amounts to detecting changes in the mean and/or variance. One algorithm is based on an adaptive (greedy) rectangular recursive partitioning scheme. The second algorithm is an optimally pruned "wedgelet" decorated dyadic partitioning. We compare the two schemes with an alternative constant variance dyadic CART (classification and regression tree) scheme which accounts only for variations in mean, and demonstrate their performance on SAR images.