We use the kernel function of the circular Fourier and radial Mellin transforms to make spatial filters and extract shift, rotation, and scale invariant features from the 2-D correlation outputs. Using a small bank of filters, this approach allows pattern recognition for multiple input objects that is invariant to their independent distortions. These Fourier-Mellin spatial filters can be phase-only and binary phase-only filters and implemented in real time using commercially available spatial light modulators. These filters are suitable as preprocessors in artificial neuron networks for adaptive and self-organizing pattern recognition. Computer simulation results are shown.