The human visual system is organized in a modular manner. This makes it
possible to analyze the different modules relatively independently from
others. The human stereopsis, for example, could be studied and
formulated in relative isolation from other modules, like distinguishing
surface orientation from texture, or shape from shading. The modules are
normally grouped into the three processes or levels of early, intermediate
and late vision. Early vision involves digital image processing tasks which
are performed relatively independently and concurrently on different
regions of the image. The result is a primal sketch which abstracts away
from the raw image consisting primarily of the geometrical distribution
and orientation of abrupt changes in the intensity function. Stereopsis is
one of several modules of intermediate vision. In order to find
corresponding points, information from late vision such as the structure
of objects, is not needed. Matching is not performed on the level of the raw
image, rather on the level of the primal sketch. Primal sketches abstract
away from the raw images and are already closer to meaning something.
The original image has a lot of redundant information. Physical and
geometrical properties of neighboring patches on a surface in the ob ject
Space normally do not vary a great deal, except along discontinuities. Thus,
much of the information of the pixels which Correspond to these patches is
redundant. The primal sketches can be thought of as a representation of
the raw image consisting of relevant information only, primarily the
geometrical distribution and orientation of intensity changes. Abrupt
intensity changes are found by convolving the raw image with the
Laplacian of a Gaussian operator, V G(x,y). The operator combines in one
Step the smoothing of the Image by taking the second derivative of the
intensity function. Where the second derivative crosses zero, the first
derivation attains its highest absolute values, and this is where the gray
levels change fastest. In the human visual system the zero crossings are
used for the matching process. Since the ZEro crossings often are identical
with the physical boundaries in the ob ject, the method of matching zero
Crossings is often referred to as feature matching.
Interestingly enough the human visual System convolves the raw image as
perceived on the retina in the same manner. Ganglion cells in the visual
cortex integrate light Intensity from photoreceptors on different
locations of the retina. Two types of ganglion cells make it possible to
transmit positive as well as negative values. In order to detect coarse and
- 604 -