High Resolution SAR Images Multi-Layer Segmentation Based on Graph Partitioning

A method based on multi-scale inherited information for SAR image segmentation is proposed. This method combines image's macroscopical and microcosmical features together, introducing the traditional single scale processing technique into the dynamic changing multi-scale analyzed framework, which makes it easy to obtain image essential features. Anisotropic diffusion equation is adopted to get multi-scale images sequences. From coarser scale to finer scale inherited graph partitioning strategy is used, for coarser scale is easy to be segmented and the segmentation results can lead finer scale segmentation. The experimental results on real high resolution SAR images demonstrate the merit of proposed method. Moreover, this method can fulfill the request of different image processing task, and consists with people's cognizing process and vision process system. The request of extracting more information from SAR images is higher than ever before, especially in the filed of remote sensing with quickly development nowadays. The pixel-based methods in image processing applications like classification, target recognition, image compression are not satisfied the requirement. The object-based methods describe the space relationship among the whole and the parts well and truly, and are capable of actualizing high-level image processing and image understanding for dealing with image objects or segments. In this paper, we propose a multi-scales and multi-level inherited segmentation algorithm which is object-based analytical method for high resolution SAR images. This method includes flowing major steps. Firstly, it uses anisotropic diffusion equation to obtain the multi-scale images. Secondly, it adopts inherited graph partitioning strategy from coarser scale to finer scale, in the same time, improves the single-scale graph partitioning to multi-scale analyzed framework. Thirdly, it establishes segmentation inherited relationships of different scales through the multi- scale graph partitioning process. At last, through all scale images the target region can be extract easily. Also, this method can fulfill the request of different SAR image analysis and processing task, and consists with people's cognizing process and vision process system. II. MULTI-SCALE ANALYSIS