A Clustering Approach Based On Spatial Location Constraints Of Pixels For Remote Sensing Image Segmentation

With the rapid development of remote sensing technology and image segmentation technology,the requirements for the collection and processing of detailed information in remote sensing images continue to increase.Image segmentation,as the most important step in image processing engineering,needs to divided images into different categories according to certain rules and characteristics,as large as possible inside the class similarity,similarity between objects,separate the categories and other categories of interest.Due to the remote sensing images are greatly influenced by environmental equipment,the image contains more details,often contain a lot of noise in the image,increases the difficulty.Clustering algorithm is one of the most widely used segmentation techniques among various image segmentation algorithms,among which fuzzy clustering algorithm remains different from traditional ones.There are only two states of belonging and not belonging to.The introduction of the concept of membership is more in line with The concept of segmentation of remote sensing images.Despite the fact that the fuzzy clustering algorithm is less complex and easier to implement,it is hard to take full advantage of the spatial position information of the pixels in images.In conclusion,such algorithm has poor anti-noise performance given its high sensitivity.This article preprocesses the image and improves the algorithm for image spatial information.In order to improve the anti-noise performance of the algorithm,the main work in this thesis as follows:Firstly,using the bilateral filter of image preprocessing,image edge details and the isolated noise filtered image gray discontinuity points,further enhance the noise resistance of the algorithm.The next step is to improved fuzzy clustering objective function,using kernel function to lower dimensional linear inseparable pixel operations mapped to high-dimensional kernel space,and by using mixed gaussian kernel parameters instead of the originally single nuclear parameters of the gaussian kernel function,improve the generalization ability of the algorithm,which improves the generalization of the algorithm.Ability.Use Mahalanobis distance instead of traditional Euclidean distance as the distance metric in the kernel space.Euclidean distance has locality,while Mahalanobis distance introduces a covariance matrix,which can better describe the global relationship between pixels and make full use of the pixel space information in the image.Finally,to further improve the spatial location constraints of pixels of the algorithm by using Markov random field model;the objective function,meanwhile,can be adjusted by using the prior probability value in the model as the correction term,give full consideration to stay segmentation pixel space restrictions associated with adjacent pixels,more fully use the image pixel space connection and restriction.The paper aims to make a comparison test among the algorithms mentioned in this paper and nine other algorithms,including FCM,KFCM,FCM＿S,FCM＿S1,FCM＿S2,En FCM,FGFCM,RFRBSFCM,BECFCM＿S and the algorithm is objectively analyzed using image segmentation accuracy(SA)and Kappa coefficient as evaluation indicators.