Case Study: Interpreting Visual Scenes - Thinking as Computation - page 122

image. Here is the background knowledge one might expect to use to help identify
these regions:
A region cannot border or be surrounded by a region of the same type.

Houses cannot be next to or surrounded by water.

Vehicles must be next to or surrounded by pavement.

Pavement cannot be completely inside any other region.

Houses, vehicles, and pavement have regular shapes (with straight edges);
grass and water have irregular shapes.

Vehicles are small; the other regions are large.

These statements constrain the permissible interpretations of regions that appear in
images in terms of size (large or small), shape (regular or irregular), borders (what is
next to what), and containment (what is surrounded by what). In reality, of course,
there are many exceptions to these rules (like houseboats that are on water, and paved
tennis courts that are surrounded by grass), but these are ignored for now.
A Prolog program that captures these facts appears in figure 6.4. It describes what is
permissible in an image of this sort. For example, the query border(pavement,grass)
returns success . This does not mean that an image must have a region of pavement
next to a region of grass but only that it could. The query border(water,vehicle) ,
on the other hand, returns failure , which means that an image cannot portray a region
of water next to a vehicle.

6.2.1 Constraints on image regions

With this background knowledge about the scene, the image can now be interpreted.
Start with a list of regions, along with the properties they have in the image, as
in figure 6.5. Approach the visual interpretation task as a constraint satisfaction

problem:
The variables are the regions of the image: R1 , R2 , R3 , R4 , R5 .

Each variable must take a value from one of the region types: grass , water ,
pavement , house , and vehicle .

The constraints that need to be satisfied are the image properties listed in
figure 6.5, in conjunction with the background knowledge about the permissible
interpretations of region types.