SAR Image Change Detection Based On Fuzzy Clustering And Deep Learning

With the development of science and technology,the exception of image precision is higher and higher.Thus,good characteristics,including high resolution,large coverage area and high interference rejection for temperature,illumination,atmosphere and weather,of the synthetic aperture radar(SAR),which make the research of SAR image change detection as a hotspot.The change detection of SAR image is a process that deal with the images acquired at different time but same place.Then the change information could be obtained.This thesis mainly researches the problems of poor robustness and low accuracy in SAR image change detection methods.The main contents are listed as follows:SAR image change detection method based on fuzzy C mean and improved bilateral filtering is presented.Firstly,the two original images are preprocessed by the Lee filter,which can remove a part of speckle noise.Secondly,the Gauss-log ratio and the neighbor log ratio are used in the proposed algorithm to produce two change images.Thirdly,the method of improved bilateral filtering is used in the proposed algorithm to fuse the two images above,and obtain a difference map.After that,the median filter is applied to the difference map is applied to the difference map in the proposed algorithm.With these,the final difference map can be obtained.The method can make full use of neighbor information and remove speckle noise effectively.Finally,an improved FCM is used to cluster the final difference map,which overcomes the defect that FCM algorithm is sensitive to noise.A method for SAR image change detection based on kernel fuzzy C mean and conditional spatial has been put forward.Firstly,the kernel function is added to FCM.This improved FCM is used to cluster difference image and then obtains a membership matrix.The kernel maps dataset to a high dimensional space through non-linear mapping method.It can also avoid the curse of dimensionality.Besides,the kernel replaces the non-robust Euclidean distance in FCM algorithm,which compensates the defect that FCM is sensitive to speckle noise.Secondly,the method uses conditional spatial method to update the membership matrix,and then obtain a new cluster center and a new membership matrix.The methodmakes full use of spatial and neighbor information of pixels to update the membership matrix again,which can obtain a more accurate membership matrix.In this chapter,the analysis and experiments show that this method can get more accurate change detection results.SAR image change detection method based on fuzzy C mean and deep learning has been presented.Firstly,joint classification and clustering are used to classify the image respectively.Thus the proposed algorithm can produce three label matrices.Secondly,some appropriate sample points are selected from the above label matrices to constitute training samples.Classification results depend on the accuracy of training samples.In order to reduce error that each method may produce,the proposed algorithm uses two methods to pre-classify the sample points.Then it selects the highest accuracy point from the sample points and combines them together.With these,the method can overcome the influence of the error to some extent.Finally,deep learning is used to detect the change.Experiments show that the proposed method can obtain lower error and higher accuracy.