is determined by the pattern or target to be recognized. To detect

the presence of a desired target while an unwanted object could appear

in the test scene, sufficient detail to discriminate the two must be

included. To pick out cars from a scene which contains both cars and

trucks, the resolution must be adequate to resolve the differences

between the two. This resolution is typically chosen in an ad hoc

fashion using the human eye to determine what resolution is required.

Computer techniques have been used to quantify the resolution

required, but the results are usually not different than what a human

would have decided by eye. Although beyond the scope of this

dissertation, the bandwidth and specific frequencies best suited to

discriminate between targets and clutter can be determined with large

computers operating on adequate training sets.

 The resolution must be adequate for target recognition. However,

oversampling beyond that resolution required will drive the CGH to

impractical limits. The resolution in the test image must match that

in the reference image yet the test image usually represents a much

larger area and larger total number of points. If the image already

exists in digital form, the resolution can be reduced by averaging the

image to produce an unaliased image of the appropriate number of

points. If an image is blurred or averaged to reduce the highest

spatial frequency, the detail above that maximum frequency is lost.

That is, all frequency components above the maximum are zero and lost.

Sampling the image properly (Nyquist criteria) permits the perfect

reconstruction of the averaged image, not the original image.

 It is worthwhile to define the concept of space-bandwidth product

(SBP) here. The bandwidth of an image is the width of the spatial