Page 402 - Introduction to Information Optics

7.5. Pattern Recognition Using Composite Filtering 387

model can then be set up for the continuous operation of the JTC. A computer
program has been written to control the joint-transform process as well as the
data association process. To initiate the coordinate, the system was first set to
run with a stationary object as input, and the position of the correlation peak
thus generated was assigned as the origin of the velocity plane coordinate. The
tracking program was then set to run autonomously with the motion sequence.
Figure 7.28b shows the actual location of the targets at each step of the
tracking cycle. A relatively simple hybrid JTC is capable of tracking multiple
targets rather accurately.

7.5. PATTERN RECOGNITION USING COMPOSITE FILTERING

Classical spatial matched filters are sensitive to rotational and scale vari-
ances. A score of approaches to developing composite-distortion-invariant
filters are available. Among them, the synthetic discriminant function (SDF)
filter (as described in Sec. 2.6.4) has played a central role for 3-D target
detection. The original idea of SDF can be viewed as a linear combination of
classical matched filters, where the coefficients of the linear combination are
designed to yield equal correlation peaks for each of the distorted patterns.
Since the dynamic range of an SDF is large, it is difficult to implement with
currently available SLMs. On the other hand, the bipolar filter has the
advantage of a limited dynamic range requirement and can be easily implemen-
ted with commercially available SLMs. Since the bipolar filter has a uniform
transmittance function, it has the advantage of being light efficient. Several
attempts have been made to construct bipolar SDF filters. However, binariz-
ation of SOF filters is not the best approach, since there is no guarantee that
the SDF will be valid. Nevertheless, iterative approaches to optimize the
bipolar SDF have been reported.
On the other hand, a simulated annealing (SA) algorithm (Sec. 2.6.5) as
applied to the design of a bipolar filter is relatively new. A bipolar filter, as we
will discuss, is in fact a spatial-domain filter, and can be directly implemented
on an input phase-modulating SLM in a JTC. To demonstrate the performance
of an (SA) bipolar composite filter (BCF), sets of out-of-plane- oriented T72
and M60 tanks, shown in Fig. 7.29a, have been used as target and antitarget
training sets. The constructed BCF, using the SA algorithm, to detect a T72
tank is shown in Fig. 7.29c. If the input scene to the JTC shown in Fig. 7.30a
is used (a BCF for detecting the T72 tank is located on the right-hand side),
the output correlation distribution can be captured, as shown in Fig. 7.30b. We
see that targets (T72 tanks) can be indeed detected from the noisy terrain,
although T72 and M60 tanks are very similar.