Page 353 - Sensors and Control Systems in Manufacturing

308
Cha p te r
S i x

image by displaying one of M remembered images. The decision
about which if any of the remembered images to display is made by
an optoelectronic analog computation of an inner product-like mea-
sure of resemblance between the input image and each of the remem-
bered images. Unlike associative memories implemented as all-
electronic neural networks, this memory does not rely on the precom-
putation and storage of an outer-product synapse matrix. Instead, the
optoelectronic equivalent of this matrix is realized by storing remem-
bered images in two separate spatial light modulators placed in tandem.
This scheme reduces the required size of the memory by an order of
magnitude.
A partial input image is binarized and displayed on a liquid-crys-
tal light valve spatial modulator which reprocesses the image in real
time by operating in an edge-enhancement mode. This preprocessing
increases the orthogonality (with respect to the inner product)
between the input image and each of the remembered images, thereby
increasing the ability of the memory to discriminate among different
images.
The light from the input image is passed through a polarizing
beam splitter, a lens, a binary diffraction grating, and another lens, to
focus an array of M replicas of the input image on one face of a liquid-
crystal-television spatial light modulator that is displaying the M
remembered images. The position of each replica of the input image
coincides with that of one of the remembered images. Light from the
array of pairs of overlapping input and remembered images is focused
by a corresponding array of lenslets onto a corresponding array of
photodetectors. The intensity of light falling on each photodetector is
proportional to the inner product between the input image and the
corresponding remembered image.
The outputs of the photodetectors are processed through opera-
tional amplifiers that respond nonlinearly to inner product level (in
effect executing analog threshold functions). The outputs of the
amplifiers drive point sources of white light, and an array of lenslets
concentrates the light from each source onto the spot occupied by
one of M remembered images displayed on another liquid-crystal-
television spatial light modulator. The light that passes through this
array is reflected by a pivoted ray of mirrors through a lens, which
focuses the output image onto a CCD television camera. The output
image consists of superpositioned remembered images, the brightest
of which are those that represent the greatest inner products (the
greatest resemblance to the input image). The television camera feeds
the output image to a control computer, which performs a threshold
computation, then feeds the images through a cathode-ray tube back
to the input liquid-crystal light valve. This completes the associative
recall loop. The loop operates iteratively until one (if any) of the
remembered images is the sole output image.