Research On Algorithms Of Remote Sensing Image Segmentation Based On Graph Theory

Remote sensing image segmentation is to segment areas of interest from the remote sensing image. Image segmentation is one of the classic problem and difficult area of digital image processing. As an important way for people to obtain information, remote sensing imaging processing has a very wide range of applications. Consequently, remote sensing image segmentation has important significance. From past to now, many types of image segmentation algorithms have been proposed, but algorithms of image segmentation based on graph theory is developed rapidly in recent years, it is a new image segmentation techniques, which use the graph theory to segment, although at the research stage, but has shown good application prospects.Currently, algorithms based on graph cut take pixels as vertexes, therefore, real-time of segmentation is poor. In order to improve it, an in-depth research on theory of graph cut has made in this paper and some theories have been applied to remote sensing image segmentation. The main research contents and research results are outlined as follows:Firstly, the background of image segmentation and graph cut are elaborated, then, Relationship and evaluative standard between image segmentation and graph partitioning are investigated, in particular, a depth research on graph theories and graph cut based on and graph spectra theory is addressed.Secondly, a remote sensing image segmentation approach is proposed, which is based on normalized cut and quarter-tree. First of all, according to gray feature, the approach use quarter-tree segmenting the remote sensing image into a large number of small partitions, then, each small partition been taken as vertex, normalized cut approach is used to segment the image among partitions from global view, by which the final segmented image can be generated. Experiment results show that the over-segmentation of quarter-tree can be eliminated effectively and computational complexity can be reduced greatly. Thirdly, this thesis suggests a remote sensing image segmentation approach, which is based on isoperimetric cut and edge growth. At first, we detected the edge of the image and then the growth of edge segment the remote sensing image into many small partitions. By Isoperimetric cut from the globe view, the indicator vector can be obtained as small areas as vertexes. Lastly, the final segmented image can be generated by clustering of vector which is used to segment the image among partitions. Comparative experiment show that the approach can effectively reduce the computation, but also the accuracy improved.Finally, on the basis of isoperimetric cut, a interactive remote sensing image segmenting approach is presented. The approach mainly consists of four steps. At the first step, we select a number of pixels at the border of object by manually then the boundary of the target area has been complete split into multiple small rectangular area as every two adjacent pixels as the apex of a rectangular. At the second step, boundary can be obtain in each rectangular by use the Isoperimetric cut to divided each rectangular into two parts. At the s third step, for every rectangular, we check whether boundary include two apex which is selected at first step, if not included, then the boundary grow until that contains the apex. The last, if we are not satisfied with the segmentation we can re-segment by back to step one. Experiments show that this method not only guarantees the accuracy of the segmentation, but also achieved fast segmentation of remote sensing image.