Synthesis of Neural Associative Memories for Artificial Vision Systems by Fuzzy Image Segmentations

A design procedure of neural associative memories to be used for robot vision systems is developed, which fits in the capabilities both of discrete-time cellular neural networks (DTCNNs) and fuzzy logic. The choice of this kind of neural networks is motivated by their architecture, suitable for storing images, and their locally connected structure, which is effective for the hardware implementation of the designed memories. In particular, fuzzy logic has been used for mapping original images into binary segmented ones, which can be stored into this kind of neural associative memory, due to the fact that the discrete-time cellular neural networks hardware realization cannot agree with the 256 gray levels of natural images and with their strongly nonlinear hystograms. The necessary storage capacity is guaranteed for the associative memory by imposing the conditions which assure the asymptotic stability for the segmented images to be memorized. The performance of the designed memory is then investigated by testing its error correction capability.