Segmentation

Everything said about greyscale images of characters holds for greyscale images in general, the extra degree of awfulness in segmentation is just as for binary images: there is less information about the extent to which things are separated and separation is not uniform in most applications. The very best thing to do with a greyscale image is to threshold it back to being a binary image if you possibly can. Unfortunately, this often gives poor quality results and more devious methods have to be used.

In Fig.2.12 we have some grey scale objects, reproduced rather unconvincingly on a laser printer. It will give you some idea of what to expect on the .tif files supplied on disk. My thanks to Miss Gek Lim for producing Fig.2.12 at very short notice.

Note that in this case we may not have a knowledge of what classes of objects are likely to turn up in the image, and that even the Trobriand Islander will have no difficulty counting seven objects, one of which is very different from the others, and two of which are overlapping. One of the six similar objects is bent, and one is short. One has the thread missing, but this is hard to see in the image on paper. This image was obtained by pointing a standard camera at the parts thrown down on a sheet of paper; although it has suffered in its many transmogrifications between camera and page, it is a higher quality image than others used in inspection in industry.

 

Figure 2.12: Nuts and bolts.

Mathematical morphology techniques can be extended to a limited extent to the greyscale case, see any of the books in the bibliography on the subject. Finding boundaries of objects by differentiating images sometimes works or can be made to work. There is no substitute for experience here, and if you do the exercises you will get plenty.