The Method Research Of SAR Image Segmentation Based On Graph Theory

In recent years,Synthetic Aperture Radar(SAR)has a wide range of applications,such as resource detection,military target identification,land management,emergency response and so on.Image segmentation is especially important in SAR image processing and interpretation,because it paves the way for the classification of objects,the detection of targets,and so on.Therefore,many SAR image segmentation methods based on different principles have been emerged.Especially,in recent years,the image segmentation method based on graph theory has aroused the concern of experts in this field because of the high flexibility and good segmentation characteristics.At present,there are a variety of image segmentation algorithms based on graph theory,but each algorithm has its limitations,it still needs the experts and scholars' research to apply these algorithms to SAR Image segmentation more efficiently.This dissertation research on the SAR image segmentation based on the Grab cut algorithm and spectral clustering algorithm.In order to improve the quality of segmentation,this dissertation mainly analyzes the principle,evolution process and implementation steps of the two algorithms,and improves the algorithm according to the difficulties and defects.The main work of this dissertation is as follows:1.A SAR image segmentation method based on Grab cut and two-dimensional entropy is proposed.The traditional Grab cut algorithm requires artificial interaction,which is not convenient enough,and the stability of the Grab cut algorithm is low due to the Gaussian mixture model(GMM),the parameters of the GMM is depended on the given initial value.Based on the above analysis,this dissertation Initialize the clusters of the SAR image by Fuzzy c-means(FCM)algorithm,according to the initial segmentation result,mark the foreground and background set,get two sets of accurate parameters of the GMM,minimize the energy function by the iterative method and separate the target area of the SAR image from the background;then the two-dimensional entropy algorithm is used to filter out the shadow of the target.Experiments with MSTAR data show that this method can achieve the segmentation of specific SAR image,the results are good and have some applicability.2.A SAR image segmentation method based on MumfordShah-G(MS-G)model and spectral clustering is proposed.The spectral clustering algorithm can converge to the global optimal solution,which is suitable for solving the actual image segmentation problem and the applicability is very good,but the traditional spectral clustering algorithm does not fully exploit the characteristic information of the image.The speckle noise in the SAR image makes the gray value of the pixel deviate from the real value,and it is not accurate enough to extract the feature directly from the original pixel.Therefore,this dissertation divides the SAR image into structural image and texture image by MS-G image decomposition model,MS-G model can maintain the uniformity and consistency of the image and the original image texture detail in the mean time.In the next step,try to use the gray value to represent the geometric structure,and use the two-dimensional entropy to represent the texture of the reaction image.Combined into two-dimensional vector based on the extracted features and construct a more accurate similarity degree matrix,then combine the spectral clustering to segment the target area of the SAR image.This dissertation experiment on the TerraSAR-X data with different objects and choice the appropriate indexes to evaluate the performance of the algorithm.