Diagonal Edges

The modified network can be expected to perform rather less well on diagonal edges than on vertical or horizontal edges because the contextual information is further away. Consider a horizontal edge, for example. Although it may have two neighbouring vertical edges that have both tripped, this information doesn't reinforce the edge because horizontal and vertical edges are connected to independent resistive grids.

The result of simulating a diagonal edge signal with both a resistive fuse network and the modified network are shown in Figure 7.4. Once again, there is a 1V signal amplitude and $\pm$0.8V noise. Clearly, the performance improvement is much less marked than it was for the vertical edge. However, whereas there are three separate gaps in the output from the resistive fuse grid, there is only one gap in the modified network. This shows that the modified network is performing its job as designed--where there are errors, then at least it is finding consistent areas. The smoothing of the voltage differential signals can't make up for a large gap, but small gaps are filled in.

Figure 7.4: Response of New and Old Networks to Diagonal Line.

There are also a large number of incorrectly segmented regions near the main feature. This can be considered to be a sign of good performance, because they do at least show where the main feature is, whereas the simple resistive fuse network has errors that may occur anywhere. Once again the light ``L'' shape has been eliminated (all images have the same noise signal). Counterbalancing this is the introduction of a dark region in the bottom left. This dark region is large enough that there it is easily classified as a region, and it is close enough to the edge that errors are likely to occur. Looking at the output intensity image, it is impossible to say that there has been much overall improvement, but equally it is clear that the performance is no worse.