Image Processing Reference
In-Depth Information
1.3 Ganglion Cells of the Retina and Receptive Fields
The ganglion cells constitute the last layer of neurons in the retina. In between the
ganglion cells and photoreceptor layer, there are four other layers of neuronal cir-
cuitry that implement electro-chemical signal processing. The processing includes
photon amplification and local neighborhood operation implementations. The net re-
sult is that ganglion cells outputs do not represent the intensity of light falling upon
photoreceptors, but they represent a signal that can be comparable to a bandpass-
filtered version of the image captured by all photoreceptors. To be precise, the output
signal of a ganglion cell responds vigorously during the entire duration of the stimu-
lus only if the light distribution on and around its closest photoreceptor corresponds
to a certain light intensity pattern.
There are several types of ganglion cells, each having its own activation pattern.
Ganglion cells are
center-surround
cells, so called because they respond only if there
is a difference between the light intensity falling on the corresponding central and the
surround photoreceptors [143]. An example pattern called (+
/
) is shown in Fig.
1.3, where the central light intensity must exceed that in the annulus around it. The
opposite ganglion cell type is (
ā
/
+), for the surround intensity must be larger than
the central intensity. The opposing patterns exist presumably because the neuronal
operations cannot implement differences that become negative.
There are ganglion cells that take inputs from different cone types in a specific
fashion that make them color sensitive. They include (
r
+
gā
)-type, reacting when the
intensities coming from the central L-cones are larger than the intensities provided
by the M-cones in the surround, and its opposite type (
r
ā
g
+), reacting when the
intensities coming from the central L-cones are smaller than the intensities provided
by the M-cones in the surround. There are approximately 125 million rods and cones,
which should be contrasted to about 1 million ganglion cells, in each eye. After a
bandpass filtering the sampling rate of a signal can be decreased (Sect. 6.2), which
in turn offers a signal theoretic justification for the decrease of the sampling rate
at the ganglion cell layer. This local comparison scheme plays a significant role in
color constancy perception, which allows humans to attach the same color label of
a certain surface seen under different light sources, e.g., daylight or indoor light.
Likewise, this helps humans to be contrast-sensitive rather than gray-sensitive at first
place, e.g., we are able to recognize the same object in different black and white
photographs despite the fact that the object surface does not have the same grayness.
The output of a ganglion cell represents the result of computations on many pho-
toreceptor cells, which can be activated by a part of the visual field. To be precise,
only a pattern within a specific region in the visual field is projected to a circular
region on the retina, which in turn steers the output of a ganglion cell. This retinal
region is called the
receptive field
of a ganglion cell. The same terminology is used
for other neurons in the brain as well, if the output of a neuron is steered by a local
region of the retina. The closest concept in computer vision is the
local image
or the
neighborhood
on which certain computations are applied in parallel. Consequently,
the information on absolute values of light intensity, available at the rod and cone
level, never leaves the eye, i.e., gray or color intensity information is not available
ā
