Image Processing Reference
In-Depth Information
o
0
o
45
o
o
o
30
o
10
5
Fig. 1.5. On the
left
, a model of the retinal topography is depicted. On the
right
, using the
same color code, a model of the topography of V1, on which the retinal cells are mapped, is
shown. Adapted after [217]
of the limited resources that the system has at its disposal, because there is a limited
amount of energy available to drive a limited number of cells that have to fit a small
physical space. Because the visual field, and hence the central vision, can be changed
mechanically and effectively, the resource-demanding analysis of images is mainly
performed in the fovea. For example, when reading these lines, the regions of interest
are shuffled in and out of the fovea through eye motions and, when necessary, by a
seamless combination of eye-head-body motions.
Half the ganglion cells in both eyes, are mapped to the V1 region. Geometrically,
the ganglion cells are on a quarter sphere, whereas V1 is more like the surface of a
pear [217], as illustrated by Fig. 1.5. This is essentially equivalent to a mathematical
deformation, modeled as a coordinate mapping. An approximation of this mapping is
discussed in Chap. 9. The net effect of this mapping is that more of the total available
resources (the cells) are devoted to the region of the central retina than the size of
the latter should command. The over-representation of the central retina is known
as
cortical magnification
. Furthermore, isoeccentricity half circles and isoazimuth
half-lines of the retina are mapped to half-lines that are approximately orthogonal.
Cortical magnification has also inspired computer vision studies to use log-polar
spatial-grids [196] to track and/or to recognize objects by robots with artificial vision
systems [20, 187, 205, 216]. The log-polar mapping is justified because it effectively
models the mapping between the retina and V1, where circles and radial half-lines
