Knowledge based geometric object recognition

The task of recognising rigid objects in an image can be greatly eased by exploiting particular geometric features of the objects. Rather than trying to match the whole object we can just match these particular features. A considerable speed advantage can be obtained over techniques like template matching, J.D. Burrows et al. (1995), as there are efficient algorithms for detecting various geometric primitives. This approach can also cope with small differences in the actual composition of the objects, as a geometric description can ignore these differences. Similarly the algorithm can be made robust with respect to clutter and occlusion/obscurement problems. We concentrate on detecting detonators in two channel X ray images. A selection of dummy detonators is presented. We exploit two particular geometric features: each detonator contains at least one dark object and we call these blobs cold spots; each detonator is approximately a cylinder which we can represent as two parallel edges lying on either side of the cold spot