424 
delivery to the multiplier/accumulator. When pipelined, the kernel for the 
lagging multiplier/accumulator is fetched from the leading one via shift 
register delay. The coefficient fast memory addresses may be selected 
randomly or sequentially via a joy stick controlled UP/DOWN counter scanner 
multiplexed into the RAM addresses. 
(iii) Image Matching Correlator/Convolver 
The sub frame processor, Fig. 1(b), is configured in the present system as a 
high speed correlator/convolver for image matching applications. It allows up 
to a 32k pixel rectangular patch of arbitrary shape and position to be 
selected from the processed (or unprocessed) image stored in a high speed 
dedicated RAM and autocorrelated, convolved, or cross correlated with a second 
image stream by means of a bank of binary correlators, Fig. 2. The 
correlation implemented is a rough approximation. The time correlation of two 
functions f,(t), £5(t) with time shift (T): 
lim 1 f 1/2 g (t) £, (t0) dt 
R(T) = Tuto 7d (2) 
is approximated by: 
Rue MP S arra] (3) 
0 k=0 
where k and j substitute for variables t and T and n is the number of pixels 
comprising the sub frame image. The output from the band of four binary 
correlator units, Fig. 2, is multiplexed into a multiplier/accumulator which 
serves to assign any prescribed weights and to sum the elemental 64 bit long 
correlations to obtain the final approximation to the correlation. A 
threshold register is present within each correlator and the 
multiplier/accumulator can be preset to any value. 
This experimental unit is designed for image sliding along the two major x, y 
window axes as a test for image matching or left-right or top-bottom pattern 
symmetry, Fig. 3. (Image sliding can be programmed to take place in any 
  
Fig. 3. An image patch is autocorrelated over a £15 pixel excursion along the 
major x, y axes of the correlation plane (a) and the converted analog 
results displayed by the dots in (b). 
arbitrary direction if required). As such it is expected that sufficient 
information can be extracted from these approximations to perform adequately