【作者】 姜丽；

【导师】 张新波；

【作者基本信息】 浙江工商大学 ， 信号与信息处理， 2010， 硕士

【摘要】 模糊C均值(FCM)聚类分析作为非监督模式识别的一个重要分支,在模式识别、数据挖掘、计算机视觉以及模糊控制等领域具有广泛的应用。该算法是一种基于目标函数的聚类算法,有着深厚的数学理论基础而且设计简单,解决问题的范围比较广,同时以该算法为基础也形成了一大批基于其他原型的FCM算法。但该算法也存在着许多亟待解决的问题,比如：需要人为定义聚类原型参数、聚类结果容易陷入局部极值点或鞍点、大数据量下算法耗时长、对特殊类型数据无法直接处理等,因此FCM聚类算法也有待于进一步改进。本文在众多的研究成果基础上,针对FCM聚类算法存在的不足,进行了较为深入的研究。基于解靴带法、样本的统计特性和核函数等理论,分别提出了两种改进的FCM聚类算法,并根据香农信息熵和模糊变差理论,提出了一种新的聚类有效性指标来评价模糊聚类的划分结果；另外,通过考虑邻域像素的影响和设计权值的方法,提出了一种可应用于含噪图像分割的加权FCM算法,同时对于较复杂的图像,还通过FCM算法的有效性函数来初始化聚类类别数,优化了分割模型；最后,将原FCM算法和神经网络算法相结合,并用于车型的分类设计中。实验结果表明,本文所改进的算法是可行的,其有效性函数也有良好的分类性能。更多还原

【Abstract】 As an important branch of unsupervised pattern recognition, fuzzy c-means cluster has wide application in the fields of pattern recognition, data mining, computer vision, fuzzy control and other fields. Fuzzy C-means (FCM) cluster algorithm is an algorithm based on objective function. It has the advantages of deep mathematical theoretical foundation, simple design as well as widely range of solving problems. Meanwhile lots of FCM algorithms basing on other prototype have been formed. However there still were many problems to be solved, such as need to define the parameters of cluster prototype artificially, cluster results easily falling into local peak, long time consuming under large volume data and unable to process the data of special types directly. So the FCM algorithm needs further improvements.On the basis of numerous research achievements, this paper does comparatively deep research on the existent insufficiency of FCM cluster algorithm. Based on the theories of bootstrapping, statistical characteristics of samples and kernel function, it proposed two kinds of modified FCM cluster algorithms respectively. It also put forward a new validity index to evaluate the partition results according to Shannon entropy and fuzzy variation theory. In addition, by considering the effect of neighbor pixels and designing weight, the paper presented a weighted FCM algorithm that could be used in the segmentation on noisy images. And for some more complex images, we initialized the number of cluster center by calculating the validation function. Finally we combined FCM algorithm with neural network and used them in the design of vehicle classification. Experiments proved that the modified algorithms were feasible and its validity function had good classification property.更多还原