Page 433 - Introduction to Information Optics

418 7. Pattern Recognition with Optics

(b)

Fig. 7.53. (a) Input patterns, (b) Initial memory matrix space with random noise, (c) Final memory
matrix representing a feature map.

IWMs and the input patterns can be written onto LCTV1 and LCTV2, and
the computer can also make decisions based on the output results of the NN.
Thus, we see that the illustrated optical NN is indeed a programmable and
adaptive network.
For demonstration, four 8x 8 pixel binary patterns (i.e., a tree, a dog, a
house, and an airplane), shown in Fig. 7.53a, are sequentially presented to the
optical NN. Figure 7.53b shows the initial memory matrix, which contains 50%
random binary pixel elements. The output pattern picked up by the CCD
camera must be normalized and the location of the maximum output intensity
can be identified by using the maxnet algorithm. The memory submatrices in
the neighborhood of the maximum output spot are adjusted based on the
adaptation rule of the Kohonon model. In other words, changes of the memory
vectors can be controlled by the learning rate. Thus, we see that the updated
memory matrix can be displayed on LCTV1 for the next iteration, and so on.
After some 100 sequential iterations, the memory matrix is converged into a
feature map, shown in Fig. 7.53c. The centers of the feature patterns are located
at (1,8), (7,1), (7,7), and (1,1), respectively, in the 8x 8 memory matrix. In this
sense, we see that the NN eventually learned these patterns after a series of
encounters. It is, in fact, similar to one human early life experience: when in
grade school, our teachers ferociously forced us to memorize the multiplication
table; at that time we did not have the vaguest idea of the axiom of
multiplication, but we learned it!

7.8.3. POLYCHROMATIC NEURAL NETWORKS

One interesting feature of optical pattern recognition is the exploitation of
spectral content. There are two major operations in NN, the learning phase and
the recognition phase. In the learning phase, the interconnection weights among
the neurons are decided by the network algorithm, which can be implemented
by a microcomputer. In the recognition phase, the NN receives an external
pattern and then iterates the interconnective operation until a match with the