In this paper we consider the wavelet decomposition of textured images with the aim to segment them. One of the key problems with using the tree-structured wavelet transform of an image is deciding how many branches of the tree are required for the accurate representation of the texture within the image, and which of these branches to use as features when clustering.
We describe here the use of two-point statistics to determine which features to select in the clustering procedure and termination of the wavelet decomposition. We present results on a set of composite Brodatz images.