Optimum use of data in space-variant optical pattern recognition

In space-variant optical pattern recognition there is a single reading which measures the correlation of the input image with the image to which a given spatial filter is matched. Here the question of optimum ways to use such readings from P different spatial filters to assign the input image to one of N categories is asked. No special relationship between P and N is necessary, and all input distortions, rotations, etc. can be taken into account.