A fast method of constructing kernel patterns for morphological associative memory

Morphological associative memories (MAMs) using kernel patterns have a lot of desirable features such as one step convergence in recalling, very large storage capacity, robustness for both erosive and dilative noise, and so on. It is, however, very difficult to decide kernel patterns when the number of training patterns to be stored is large and when training patterns have no unique feature bits. We derive a novel method of constructing kernel patterns for MAMs by examining the relation between characteristics of kernel patterns and outputs of MAMs. Computer simulation results show that the proposed method is more than 10 million times faster than the conventional method using trial and error. Moreover, since the proposed method can construct kernel patterns from original training patterns, it keeps remarkable features of the MAMs.