Research On Noise Tolerance Of Morphological Associative Memory Networks

Compared with traditional associative memory, Morphological Associative Memories (MAM) has many attractive advantages, such as unlimited storage capacity for auto-associative memories, MAM not only has capability of handling binary pattern but also can deal with the real mode, as well good noise-tolerance to either erosive or dilative noise and one-shot recall, there are, however some problems, such as it has a good performance of single noise-tolerating; to the contrary the mixed noise is ineffective. The model can only abide by one to one association, not one to many associations, so it leaves us much room to research. Therefore, from the perspective of the dual kernel and modularity, this paper mainly has focused on how to improve the noise-tolerance of Morphological Associative Memories.In the past, In order to improve the performance of noise-tolerance, the kernel method has been put forward, although received a certain effect, but it's difficult of finding the kernel. So it also proposed the idea of modular, it tries to improve the capacity of noise-tolerance by the Inhomogeneity of the noise distribution. Despite the shortage of detailed algorithm and steps to realize, the idea is acceptable. Combining the associative memory W robust in recalling patterns that are distorted due to erosive changes and the associative memory M robust in recalling patterns that are distorted due to dilative changes, this paper researches and improves the dual kernel, The dual kernel can be put together with the kernel to learn from other's strong points to offset ones'own weakness and to make the effect of hetero-associative memories and pattern recognition better. In this paper, the necessary conditions of the dual kernel need to be meeting have been defined clearly, its properties have been researched, and a way to accelerate the search for dual kernel has been found. The dual kernel concept has been further extended in the scale space and its range of application expanded then its usability enhanced. Through simulation experiments of the character and true-color images, the dual kernel has a good performance of tolerating random noises. In addition, it's researched that how to improve the capable of noise-tolerance using modularity. The segmentation and integration methods of the patterns, such as the row-column partitioning and checker board partitioning are proposed. The construction of the associative memory at the Learning phase and the process of the recall at the recalling phase of row-column partitioning and checker board partitioning is described meticulously. Moreover, Combining row-column partitioning and the voting matrix, the row-column modularity associative memory model based on the voting matrix has been put forward. In this paper its treatment process has been described in details. The method of modularity separates a large pattern into small ones, avoiding this verbose process of finding the kernel and dual kernel. In the stage of memory, the capacity of noise-tolerance has been improved though index location. Its feasibility and validity has been confirmed by the simulation results. In order to accelerate the process of modular associative memory, the paper puts forward the parallel algorithm of Modular associative memory based on Cannon method, its design and implementation has been proposed.