Curvature analysis

While the primary operation of vision has to be, without doubt, the detection of local edges in a scene, it has to be the recognition & identification of local objects which are of utmost importance at the perceptual level. While there have been huge amounts of experimentation concerning the inter-relationships of various parameters for the detection task (see, for instance, the early Chapters of ‘Vision and Acquisition’), significant experimentation on the effects of such various parameters on the higher level tasks is much more difficult to control. A major point of contention is what exactly are the properties of an object which control its recognisability.

I have argued, after very considerable research study & experimentation, that whatever else might have an influence on recognition, one thing which seems to have an inescapable influence is the shape of an object - that is, the existence & geometric association of corners and zones of greater or lesser curvature.

It has been shown with the aid of SHV that, given the demonstrable ease with which sub-pixel local edge location & orientation can be derived by what has been determined to be the basic early processing carried out by human vision, it is an almost trivial extra task to derive, for each pixel along a detected edge, the local curvature at that point. Then from this it is again almost trivial to establish curvature trends around any continuous profile or part profile.