Morphological Decomposing Associative Memory Based On Dynamic Kernel And Its Robust Analysis For Random Noise

Associative memory is one important function of human brain cells, and one important aspect of artificial neural network is to simulate the associative memory function of human beings. The morphological associative memories proposed by G.X.Ritter et al at the end of the 20th century is an effective way to overcome the drawbacks classical associative memories, such as the limited sample storage capacity and necessity of iterative memory. In addition, the network has adequate anti-noise capability in processing binary images with single noise, but it seems inadequate in dealing with random noise. Morphological associative memories and its improved versions cannot effectively deal with the issue of ambiguous interpretation, thus B.Kosko proposes fuzzy associative memories, which solves the issue of ambiguous interpretation but has quite poor storage capacity.Papers by combining morphology and fuzzy associative memories, propose fuzzy morphological associative memories. This new network solves the problem of poor storage capacity and also has a strong anti-noise capability against single noise, but can hardly suppress any random noise, which limits its usage.Four aspects are mainly made in this paper.The first aspect is that we arrived at a new algorithm by combing dynamic kernel with images decomposition algorithm.And the results show the algorithm has a strong ability to resist random noise and shortcomings of the algorithm are described at the same time.The second aspect is that the Morphological decomposing associative memory arithmetic of dynamic kernel are applied to MBAM, FMBAM andαβassociative memory networks, making these three-networks have ability to resist random noise, and the application scales are promoted to gray-scale images and color images. The Third aspect is we come up with parallel decomposing associative algorithms which shorten greatly associative memory time。Finally, we succeeded in applying this algorithm to the image recognition and verified the above conclusions by a large number of simulation tests.