A Morphological Associative Memory Employing A Stored Pattern Independent Kernel Image and Its Hardware Model

An associative memory provides a convenient way for pattern retrieval and restoration, which has an important role for handling data distorted with noise. As an effective associative memory, we paid attention to a morphological associative memory (MAM) proposed by Ritter. The model is superior to ordinary associative memory models in terms of calculation amount, memory capacity, and perfect recall rate. However, in general, the kernel design becomes difficult as the stored pattern increases because the kernel uses a part of each stored pattern. In this paper, we propose a stored pattern independent kernel design method for the MAM and design the MAM employing the proposed kernel design with a standard digital manner in parallel architecture for acceleration. We confirm the validity of the proposed kernel design method by auto- and hetero-association experiments and investigate the efficiency of the hardware acceleration. A high-speed operation (more than 150 times in comparison with software execution) is achieved in the custom hardware. The proposed model works as an intelligent pre-processor for the Brain-Inspired Systems (Brain-IS) working in real world.