SAR Image Segmentation And Change Detection Based On Fuzzy Clustering And Key Pixels

Nowadays,synthetic aperture radar(SAR)images are very common and important in military and civil areas.The imaging process of SAR image is not affected by the time,light and the weather,which makes this kind of image provide more useful information.However,because of the special imaging mechanism,SAR image itself contains many speckle noises,and this kind of noise is multiplicative,which makes the processing of the SAR image very challenging.SAR image segmentation is a very essential part of the domain of SAR image processing,which is the basic work of subsequent SAR image interpretation.And SAR image change detection is to detect the changed regions from two SAR images obtained from the same place but in different times,which also has very important applications in military and civil areas.This paper focuses on the problems of the low robustness to the speckle noise and low accuracy of the existing SAR image segmentation and change detection methods.The main contents are as follows:An unsupervised method based on fuzzy clustering and key pixels is proposed for synthetic aperture radar image segmentation.The method only performs fuzzy clustering on some key pixels and uses the clustering result to quickly segment the remaining pixels,which avoids involving all pixels in the time-consuming clustering.Specifically,first,some pixels are selected from the original SAR image according to a certain rule.These pixels which carry the basic information of the whole image are called key pixels.Then,a fuzzy clustering method based on spatial information is used only for segmenting key pixels.A very accurate clustering result can be obtained quickly because of the small-amount key pixels and the using of spatial information.Finally,based on the clustering results and a similarity metric rule which is robust to speckle noise,labels of the non-key pixel are determined,and the whole image can be segmented quickly and accurately.This method only uses a few pixels to execute time-consuming clustering,so accurate result can be obtained in a short time.At the same time,the spatial information is fully used to further enhance the robustness of the method to speckle noise.A segmentation method based on improved Turbo super pixels and key pixels is proposed for SAR image.The proposed method improves the accuracy of Turbo super pixels' extraction by fusing multiple edge detection results.Meanwhile,the key pixels and super pixels are combined together to provide more effective information for SAR image segmentation process in order to improve the segmentation accuracy.Specifically,first,Turbo super pixels are obtained based on the edge information by fusing the responses of three edge detectors;secondly,the key pixels are selected according to the local maximum pixel rule,and the clustering process is executed only on the selected key pixels;thirdly,Turbo super pixels are segmented based on the distribution and the clustering results of key pixels;finally,a filtering process for super pixels is executed to further enhance the accuracy of segmentation.The method takes advantage of the information of super pixels and key pixels,which improve the robustness of the segmentation to the speckle noises.An unsupervised method based on fuzzy clustering and self-paced learning is proposed to deal with SAR images change detection problems.This method detects changed regions from the difference image by constructing a classifier: first,after pre-classifying the difference image,some samples are selected based on the uniform selection strategy to avoid the number of samples in one certain class being too small and to ensure the diversity of samples;secondly,the self-paced learning framework is introduced to train the classifier.The cost function is applied to judge the difficulty of one sample,and the difficulty of the sample is gradually increased in the training process,which makes the learning more and more difficult and ensures the stability of learning result;finally,the classification results and spatial neighborhood are combined together to ultimately determine the label of each pixel.This method benefits from the self-paced learning to get a classifier with high robustness to noises and to improve the final accuracy.