【作者】 邓赵红；

【导师】 王士同；

【作者基本信息】 江南大学 ， 轻工信息技术与工程， 2008， 博士

【摘要】 模式识别技术是人工智能的重要研究内容。基于各种技术,几十年来各种不同的模式识别方法得到了广泛的研究。其中,随着神经模糊技术的发展,基于神经模糊技术的模式识别方法也得到了长足的发展,引起了众多学者的广泛关注,形成一个独特的研究方向:神经模糊模式识别技术。目前,基于神经模糊技术的模式识别及其相关技术已经得到了较深入的研究,一些成果已成功高效地应用于不同的领域。虽然如此,该类技术依然面临着许多重大的挑战。其中几个关键的挑战是:1)如何构建更鲁棒的神经模糊模式识别算法。2)如何开发基于新模型的神经模糊模式识技术。3)如何把神经模糊模式识别及其相关技术应用于更广泛的领域,如生物信息学、计算机视觉等。针对上述的挑战,本课题进行了相关的研究。所研究内容主要涉及三个部分,分述如下。第一部分,包含第二章到第五章,主要探讨了鲁棒的神经模糊模式识别技术。具体地,第二章针对模糊聚类神经网络FCNN对例外点敏感的缺陷,通过引入Vapnik’sε?不敏感损失函数,重新构造网络的目标函数,并根据拉格朗日优化理论推导出新的鲁棒模糊聚类神经网络及其算法RFCNN。第三章针对极大熵聚类算法MEC对例外点较敏感和不能标识例外点之缺陷,提出了一种鲁棒的极大熵聚类算法RMEC。第四章提出了一种较鲁棒的基于视觉原理和WEBER定律的TSK模糊系统建模方法。第五章提出了一种新的级联MLP神经网络CATSMLP。证明了CATSMLP神经网络等价于一种特殊的基于演绎模糊推理的级联模糊推理系统CATSFIS;由于级联模糊逻辑推理较之于if-then模糊逻辑推理对噪声的干扰具有较小的误差上界,从而推导出CATSMLP神经网络较ATSMLP具有更好的鲁棒性。第二部分,包含第六章到第八章,主要探讨了基于球模型的神经模糊模式识别技术。具体地,第六章提出了一种基于核化技术的模糊核超球感知器分类算法,该算法通过核化技术把样本数据映射到高维特征空间,并利用超球感知器学习寻找高维特征空间的决策超球,从而得到各类样本的决策函数。第七章基于最小最大概率策略和模糊技术提出了一种新的分类学习机:模糊超椭球学习机MPFHM。第八章探讨了压缩集密度估计器RSDE和最小包含球MEB之间的关系,证明了RSDE能被视为一个特殊的MEB问题。进一步引入基于核集的最小包含球逼近策略开发出了快速的压缩集密度估计器FRSDE,并有效地应用于分类、建模及图像分割。第三部分,即第九章,基于模糊推理规则提出了一种具有自适应学习功能的自动弹性图像配准方法。进一步地,把形变视频跟踪看作一个动态图像配准问题,提出的弹性配准方法被应用于视频跟踪。更多还原

【Abstract】 Patten recognition is one of important tasks of artificial intelligence research, which has been extensively studied in the past tens of years. With the development of neural-fuzzy techniques, the neural-fuzzy techniques based pattern recognition techniques attract more and more attentions of the researchers and then a new research topic, i.e, the neural-fuzzy pattern reconition has emerged. Nowadays, a lot of important advancements have been achieved. However, the neural-fuzzy pattern recognition still confronts many challenges. Among of these challenges, several crucial challenges can be described as follows: 1)how to develop more robust neural-fuzzy pattern recognition algorithms; 2)how to develop the new-model based neural-fuzzy pattern recognition techniques; 3)how to apply the neural-fuzzy pattern recognition techniques to more extensive research fields, such as bioinformactis, computer vision and so on.Motivated by the above challenges, several issues are addressed in this study, which mainly involves the following three parts.In the first part, the robust neural-fuzzy pattern recognition techniques are investigated, which contains Chapter 2 - 5. In Chapter 2, in order to overcome the weakness of sensitivity to outliers of fuzzy cluster neural networks(FCNN), a robust fuzzy clusting neural networks algorithm RFCNN is presented. In Chapter 3, a novel robust maximum entropy clustering algorithm RMEC, as the improved version of the maximum entropy clustering algorithm MEC, is presented to overcome its drawbacks: very sensitive to outliers and uneasy to label them. In Chapter 4, the TSK fuzzy system modeling is re-considered from a new point of view and a more robust TSK fuzzy system modeling approach based on the visual-system principle and the Weber law is presented. In Chapter 5, we present a new MLP model called cascaded ATSMLP (CATSMLP) where the ATSMLPs are organized in a cascaded way. The proposed CATSMLP is proved to be functionally equivalent to a fuzzy inference system based on syllogistic fuzzy reasoning. Meanwhile, we in an indirect way indicate that the CATSMLP is more robust than the ATSMLP in an upper bound sense.In the second part, the ball-model based neural-fuzzy pattern recognition techniques are investigated, which contains Chapter 6 - 8. In Chapter 6, a novel fuzzy kernel hyper-ball perceptron is presented to realize the classification desicion. In Chapter 7, a novel classification machine called the minimax-probability based fuzzy hyper-ellipsoid machine MPFHM is proposed using the hyper-ellipsoid with the minimax probability principle and fuzzy concept. In Chapter 8, in order to overcome the shortcoming of the high time and space complexities of reduced set density estimator RSDE, a fast reduced set density estimator algorithm FRSDE is proposed. The finding that RSDE is equivalent to a special MEB problem is derived and with this finding the fast core-set-based MEB approximation algorithm is introduced to develop the algorithm FRSDE.In the third part, which contains Chapter 9, the applications of neural-fuzzy pattern recognition techniques in other research fields are investigated. In Chapter 9, in terms of the characteristics of elastic image registration, a fuzzy-inference-rule based flexible model is proposed for the automatic elastic image registration. Furthermore, we apply the proposed registration algorithm to visual tracking.更多还原