We also develop a way of objectively evaluating texture segmentation algorithms on natural and synthetic texture patches. Finally, our multiscale segmentation approach is demonstrated on several families of real-world images. It is shown that quality of the segmented results at the different scales is significantly improved.

A novel method of performing multiscale segmentation of images using texture properties is introduced. Various methods of segmentation at a single scale, including texture segmentation, are described and compared with the K-Means clustering of texture vectors used in this thesis.

We survey the state of the art in multiscale segmentation and identify some drawbacks in the traditional approach to multiscale segmentation---image pyramids and quad-tree decomposition, especially in typical applications that make use of the segmented results. We then introduce the idea of multiscale segmentation within the context of the segmented regions themselves instead of, as is traditional, working in the context of the original image. It is shown that this new multiscale approach can be incorporated into the K-Means clustering technique as a steady relaxation of inter-cluster distances.