Methodology

Humans tend to concentrate on sharp edges rather than any other kind of signal [13]. These edges are not differences between individual pixels, but long sequences of such differences. An isolated pixel can be determined not to be a real feature, because it does not form part of a sequence like this.

In order to make the circuit more intelligent, it was decided to make the circuit follow continuous sequences of differentials. This involves abstracting the entities the circuit operates on from individual pixels to edge elements, that is, differences between two neighbouring pixels.

A number of assumptions were made before embarking on the design of this network:

• The purpose of the network is to detect the edges that define large regions that represent a single ``object'' in an image.
• Because an ``object'' could be any kind of complicated shape, for this project an ``object'' is limited to a large region in which there is a consistent or smoothly varying intensity.
• Objects that are only defined by straight lines are not objects of interest. This is important simply because it helps the design of this network. Most real-world images do not have important regions defined in this way. An important exception is line drawings. This network will not be appropriate for use on these images.

An example of the kind of regions that should be detected, drawn manually, is shown in Figure 6.1. There is no attempt to detect the whiskers, for example, because although they are strong features, they do not in themselves define regions. However, a region which includes whiskers, such as that in the bottom of the image, is different to a region without, such as the upper lip.

Figure 6.1: Image of Mandrill with Ideal Segmentation.