Page 434 - Introduction to Information Optics

7.8. Neural Pattern Recognition 419

stored pattern is obtained. The iterative equation for a two-dimensional NN is
given by

where U lk and l/, ; represent the 2D pattern vectors, T lkij is a 4D IWM, and /
denotes a nonlinear operation, which is usually a sigmoid function for
gray-level patterns and a thresholding operator for binary patterns.
A polychromatic neural network (PNN) is shown in Fig. 7.54, where two
LCTVs are tightly cascaded for displaying the input pattern and the IWM,
respectively. To avoid the moire' fringes resulting from the LCTVs, a fine-layer
diffuser (e.g., Scotch tape) is inserted between them. To match the physical size
of the IWM, the input pattern is enlarged so the input pattern pixel is the same
size as the submatrix of the IWM. This is illustrated in Fig. 7.55. The
summation of the input pattern pixels with the IWM submatrices can be
obtained with a lenslet array by imaging the transmitted submatrices on the
CCD array detector. By properly thresholding the array of detected signals, the
result can be fed back to LCTV1 for the next iteration, and so on.
The liquid crystal panels we used happen to be color LCTVs, for which the
color pixel distribution is depicted in Fig. 7.56. Every third neighboring RGB
pixel element is normally addressed as one pattern pixel, called a triad.
Although each pixel element transmits primary colors, a wide spectral content
can be produced within each triad. If we denote the light intensity of the pixel
element within a triad as I R(x,y\ I G (x,y), and 7 B(x, y), the color image
intensity produced by the LCTV is

/(.x, y) , y) + ] G(x, y) + / B(x, y). (7.76)

White
Light
Source

Fig. 7.54. A polychromatic NN using cascaded color LCTVs.