Unsupervised texture segmentation using Hermite transform filters

Abstract

A Hermite transform multi-channel filtering approach to unsupervised texture segmentation is presented. Texture feature images are obtained by first applying Hermite transform filters of various orders and in two directions to the input images, then next passing filtered components through a nonlinearity and computing local averages. These texture feature images are decimated in a hierarchical pyramid image representation. Segmentation by a square error clustering technique is performed on the decimated feature image with the classification being propagated down the pyramid to obtain pixel labels of the segmented image. Good segmentation results were obtained with about ten channels.