Spot counting

One of the more fundamental things to be done in pattern recognition is simply counting the objects in an image. Going back to Fig.2.10, the paper clips, we see that the objects may overlap, and that this can very effectively clobber segmentation methods. If the objects are simple in shape, say disk shaped, then something can be done, but it is mostly nasty and ad hoc and therefore gives a thrill only to debauched computer programmers.

Some of the images on disk are microphotographs of cells, and it is frequently necessary to count the cells of a given type or colour. They usually overlap. Erosion of the cells until they fall into separate points, boundary smoothing and colour quantisation have been used by Dr. Chris deSilva and his students at the Centre for Intelligent Information Processing at the University of Western Australia, to obtain reliable cell counts. More advanced methods will be discussed later. It should be noted that there are an awful lot of cells to count, and automating the counting process is a critical bottleneck of mass screening schemes, in, for example, the early detection of breast cancer.