Research On Robust Fuzzy Clustering Algorithm Based On Feature Selection

Image segmentation divides images into specific regions and extracts interest objects.It is the key step and foundation to deal with complex tasks such as image understanding and target recognition.At present,image segmentation technology is widely used in the fields of computer science,pattern recognition,machine learning,space communication,medical diagnosis and other fields.The new achievements of image segmentation have greatly promoted the development of related fields.The essence of image segmentation is pixel clustering.The image information itself has fuzzy uncertainty because of the imaging mechanism,transmission noise interference and visual perception.Therefore,the application of fuzzy clustering technology to solve the structure discomfort problem of image segmentation has been widely concerned by many scholars.The fuzzy C-mean clustering algorithm(FCM)is used as an unsupervised clustering algorithm.It uses the square Euclidean distance to describe the difference between the sample and the cluster center,which is only suitable for clustering of data samples with different samples and variance.For this reason,the Gustafson-Kessel algorithm uses Mahalanobis distance as a similarity measure to replace the Euclidean distance in FCM algorithm in order to adapt to the non spherical structure problem.Considering the influence of key features on clustering,the fuzzy clustering algorithm based on the feature selector(FSFCM)overcome the sensitivity of the FCM algorithm,which is more suitable for the complex multi-scale target segmentation.The neighborhood local fuzzy C-mean segmentation algorithm(FLICM)takes account of the influence of neighborhood spatial information on clustering and improves the noise immunity of FCM algorithm.The FCM algorithm,FSFCM algorithm and FLICM algorithm are improved by the concept of Markov distance,Markov and feature selection.The main contents of this paper are as follows:Color images are rich in information,and the algorithm of using only the segmentation of gray images can not reflect the relationship between spatial features of color images.Firstly,the square Euclidean distance in the algorithm of FCM is replaced by the Mahalanobis distance with covariance,so that effectively utilize the algorithm of FCM covariance information of the cluster.Secondly,combining the thought of the algorithm of FCM-S2,the neighborhood spatial information is embedded in the target function,and the noise resistance performance of the algorithm is improved.Finally,a robust clustering segmentation algorithm suitable for color images is obtained.Through the segmentation of multiple standard color images under different noises by the improved algorithm and a variety of classical algorithms.The experimental results show that the improved algorithm is improved in the segmentation performance and the robust performance of the noise.In order to improve the anti noise ability of the algorithm of FSFCM,a robust fuzzyclustering algorithm for adaptive feature selection is proposed in this paper.Firstly,the priori probability of Markov random field is established considering the correlation of pixel neighborhood space.Secondly,the classification membership degree and the prior probability in the pixel neighborhood were combined,the combination of the exponential power function was used as a smoothing noise factor to constrain the prior probability through the KL divergence.The constraint was transformed into a regular factor and embedded into the algorithm of FSFCM.Comparing the experimental results of segmentation of the noisy panchromatic remote sensing image and the simulated image by the improved algorithm and a variety of algorithms.it is proved that the segmentation results of the improved algorithm have good consistency and strong anti noise robustness,and meet the needs of the noise image segmentation.algorithm of FLICM can not take account of the influence of different features on clustering results.a improved method based on the algorithm of FLICM is proposed in the fifth chapter of this paper.First,we add the constraint of the membership degree to the space distance in the local information function.Secondly,the feature salience is introduced into the objective function.According to the Lagrange multiplier method,the optimal expression is solved for the objective function,and the neighborhood weighted trust is added to the classification membership representation.finally,a local fuzzy clustering segmentation algorithm based on feature selection is obtained.comparing various algorithms with the improved algorithm,the improved algorithm can obviously reduce the iteration time on the basis of guaranteeing the quality of segmentation.