Multiscale relaxation labeling of fractal images

This paper describes the application of multi-scale relaxation to automatically detect pavement distress. Pavement distress detection is a di cult task which simple edge detection schemes perform poorly. We have chosen to use relaxation labeling to improve upon an initial edge-based segmentation. This work is based upon a fractal model of pavement distress. The scale-invariance property of fractals suggests that information at di erent scales of resolution may be combined to improve segmentation. Thus, we have developed a multi-scale relaxation technique for use in a pavement distress detection system. Straightforward linear interactions fail to capture the complexity of pixel interactions for this problem. To better model pixel interactions, we have included non-linear terms in the relaxation process. Symmetry arguments and careful engineering allow a 93% reduction in the complexity of this approach. To demonstrate the necessity of the multi-scale approach, examples with and without multi-scale relaxation are shown. We found that performance was greatly improved by multi-scale relaxation.