Research On Possibilistic Fuzzy Clustering Algorithm For Multiple Kernels

As an important branch of the clustering field,the validity of the fuzzy C-means clustering algorithm was largely limited to intra class compact,good separation between classes and globular clusters.It can solve this problem by mapping non-linear data to the appropriate high-dimensional feature space,but in the high-dimensional feature space,the choice of nuclear kernel is sometimes complicated.In this dissertation,we propose a fuzzy C-means clustering algorithm for multiple kernel probabilities.This algorithm can effectively solve the shortcomings of fuzzy C-means clustering algorithm by automatically adjusting the kernel weight and the multiple kernel way by using the advantages of fuzzy clustering method and probabilistic clustering algorithm.Furthermore,the algorithm can avoid the uncertainty of the selection of kernel functions and increase the anti-transformability of the algorithm.For the selected multiple kernel function,the combination of weights can satisfy different data or image for various kernel functions of the preferences of the demand to calculate the best match the weight value.In the absence of any prior circumstances,not only it can be accurately divided,but also can be divided into non-linear cluster samples.The experimental and simulation results show that the algorithm proposed by us is excellent in terms of the correctness of the data classification and the comparison of the index,and has high practicability.The main work and innovation are as follows;(1)A brief introduction is provided for several classical fuzzy clustering algorithm methods.Such as the fuzzy C-means clustering algorithm(FCM),the possibilistic C-means clustering algorithm(PCM),the possibilistic fuzzy C-means clustering algorithm(PFCM),the kernel fuzzy C-means clustering algorithm(KFCM)and so on.This dissertation expounds the research status of the improved FCM algorithm,and analyzes the corresponding problems,such as the problem of sensitive points to the outliers and noise points,the problem of low accuracy of the classification of high dimensional space,easy to produce the problem of consistency clustering,the choice of kernel function.(2)A fuzzy C-means clustering algorithm for multiple kernels is proposed.The algorithm can effectively use the advantages of fuzzy clustering method and possibilistic clustering algorithm,and avoid the problem of which the fuzzy C-means clustering algorithm is more sensitive to noise points and which the possibilistic clustering algorithm is easy to produce consistent clustering.Furthermore,the algorithm can avoid the uncertainty of the selection of kernel functions by the ordinary algorithm.By making the non-linear data operation,the linear operation of the data in the normal data can be mapped to the high-dimensional data space,which increases the antiTransformability.The newly proposed algorithm is applied to the UCI data sets and the images to compare the advantages of the proposed algorithm and the classical algorithm of the past.It fully compares the membership degree,the clustering center and the classification accuracy.(3)An Enhanced multiple kernel possibilistic fuzzy C-means clustering(EMKPFC)algorithm for image segmentation is proposed.The existing image segmentation algorithm based on fuzzy clustering is sensitive to noise and cannot properly deal with the relationship between gray features of images and neighborhood pixels.In order to avoid the uncertainty of the selection of kernel function,the algorithm can avoid the anti-transformability of the algorithm.For the selected kernel function,the combination of weights can satisfy the demand of different images for various kernel functions,and calculate the best match weight value.The experimental results of artificial images,real images and medical images show that the proposed algorithm has better effect than other related image clustering algorithms based on fuzzy clustering.