Boundary Cue Operators for Model-based Image Processing

An efficient method of using explicit shape models of objects in boundary instantiation is to apply one dimensional edge searches in locations where boundaries are likely to occur. In many important cases, linear edge operators produce at best only weak responses. We investigate here the use of three different statistical measures applied over a sliding 'dipole' as candidates for detecting weak boundaries. Their performance is compared with an implementation of the Canny operator as a benchmark on synthetic images of step edges in random noise and on certain difficult real images. In the former case their performance compares favourably with the Canny operator, while in the latter case they can produce significant responses where the Canny operator detects only weakly or not at all.