Scale-space pattern processors: are they robust to noise and occlusion?

An emerging interest in the field of computer vision and pattern recognition has been that of a `scale-space´ in which an image is progressively simplified in a manner that does not introduce artefacts. The idea is useful as it allows the input pattern space to be sampled at an appropriate scale and hence reduce data rates without losing important features. But how are such systems affected by noise or occlusion? In this paper we discuss the performance of the conventional linear diffusion processor and compare it to a class of morphological systems. We show, by using very stylised targets in both synthetic and real images, that diffusion-based systems are sensitive to noise and occlusion. The morphological systems we study have performance that is as good as, or better than, the benchmark diffusion system