【作者】 杨群生；

【导师】 余英林；

【作者基本信息】 华南理工大学 ， 通信与电子系统， 2000， 博士

【摘要】 模糊理论和技术是当今学术界的研究热点，本文围绕模糊神经网络和模糊图象处理展开研究，分别对模糊联想记忆(FAM)神经网络和模糊图象噪声消除进行了细致深入的探讨。本论文的主要工作如下： 1)对于Bart Kosko提出的max-min FAM网络模型，针对其学习算法不能确保可靠地存储多个模糊模式对的缺陷，提出了一种新的学习算法——多模糊模式对FAM连接权矩阵的神经网络学习算法，并且给出了相应的理论结果和严格的数学证明，同时通过实验验证了该算法的有效性。该算法成功地解决了多模式对的存储问题，利用该算法可以用尽可能少的连接权矩阵来存储给定的模糊模式对，有效地减少存储空间，提高存储容量，而且容易实现。对于模糊双向联想记忆网络，也给出了连接权矩阵的神经网络学习算法，并讨论了它的容错性。 2)将上述关于max-min神经网络模型的有效算法推广到基于一类最大T-模运算的FAM神经网络模型，并且给出了相应的理论结果和严格的数学证明。因为FAM网络内部算子不同，FAM系统的用途和效果也不同，所以该算法大大地扩展了FAM的适用范围。 3)在文献[122]的基础上，设计了一种新的滤波器--模糊检测加权均值(Fuzzy Detection Weighted Mean，简记为FDWM)滤波器。首先以盐椒噪声图象的直方图为基础，建立模糊隶属函数，然后进行了算法设计，结合新的检测算法进行噪声消除，提出了一种非常有效的盐椒噪声消除算法—FDWM算法，并分析了该算法的原理，而且做了大量的实验。实验结果表明，FDWM滤波器的去噪效果远远优于常规算法，而且性能稳定，适合于处理各种噪声率的盐椒噪声图象。特别是当噪声率超过50％时，FDWM滤波器的优越性更加突出，无论其主观视觉效果还是其峰值信噪比或均方误差都表明了这一点。 4)在上述FDWM算法的基础上，提出了基于模糊中值滤波(Fuzzy MedianFilter，FDF)的FDWM滤波器—FDF_FDWM滤波器。虽然上述FDWM滤波器可以有效地消除盐椒噪声，但它不适合于消除图象的随机噪声。为此，本文首先提出迭代模糊中值滤波(Iterative Fuzzy Median Filter，IFMF)算法，然后利用IFMF更多还原

【Abstract】 The fuzzy theory and fuzzy technique are the current focus of academia at present. Around the fuzzy neural networks and fuzzy image processing, Fuzzy Associative Memory (FAM) neural networks and the removal of noise in image are studied, some new algorithms are proposed and some new implementation algorithms are proposed in this dissertation. The main works are listed as follows:1) Because the max-min composition fuzzy Hebb Fuzzy Associative Memories (FAM) networks model proposed by Bart Kosko can’t ensure multiple fuzzy pattern pairs to be encoded in a FAM associated weight matrix, a new neural networks learning algorithms for multiple-pattem pairs connection weight matrix of FAM and its relevant theoretical results and strict mathematics proofs are presented. At the same time, its effectiveness is testified by simulations. The algorithm is used to solve the problem of storage for multiple fuzzy pattern pairs successfully. Multiple fuzzy pattern pairs can be encoded to store in associated weight matrixes of FAM as few as possible by the algorithm, so it can cut down storage space and improves its storage capacity greatly, moreover this algorithm can easily be implemented. For the bi-directional FAM, a kind of neural network learning algorithms for associated weight matrix is presented and its fault-tolerances is discussed too.2) Generalize the algorithm of max-min composition FAM aforementioned into max-T norm operations FAM. Neural networks learning algorithm for associated weight matrix of fuzzy associative memory which is based on a class of T -norm operations and its theoretic results and strict mathematics proofs are presented. If the interior operator of FAM networks is different, the different purposes and effects of FAM system can be obtained, so the algorithm extends the application of FAM greatly.3) Based on literature [122], a new filter—Fuzzy Detection Weighted Mean (FDWM) filter is presented in this dissertation. According to the histogram of corrupted image and its statistic characteristic, first we construct the fuzzy membership functions更多还原